{ "format": "neural-download-model-family-v1", "id": "qwen-flash-next", "collapse_coverage_contracts": true, "primary_packet_id": "qwen38-flash-next-fp8-tp4-research", "name": "Qwen Flash-Next", "display_name": "Qwen3.8 Flash-Next · 125B-A6B", "publisher": "Qwen", "updated_at": "2026-09-07", "summary": "Qwen's experimental 125B-A6B hybrid-attention MoE. The official FP8 export serves on four B70s with selective host placement, and all 25 practical TP4 eager-text MTP/context cells through 8K are classified. The retained current runtime now independently qualifies TP4 eager MTP0 at short context and exact active 4K: 5.224 tok/s on the established short screen and 4.758 tok/s conventional at exact 4K, with 6/7 semantic quality, 16/16 repeats, an exact cache-zero 4K needle, and card-clean teardown. At 16K, current-source MTP0 completed one exact generic-depth request, then the semantic program produced one correct fresh response, a corrupted same-server repeat, and a separate fresh-server worker timeout at 1,600 computed prompt tokens. This is Grade-D quarantined capability and nondeterministic runtime-stability evidence; it does not authorize 24K/32K. MTP2's bounded 16K treatment tranche is exhausted. MTP3 remains the preferred exact-4K recipe at 15.502 tok/s decode. The target-only official-thinking MTP0 profile passed 25/25, and graph, vision, other topologies, clean-host replay, and deployment qualification remain open.", "featured_results": [ { "role": "hero", "label": "Qwen3.8 Flash-Next FP8 · TP4 lossless MTP1, exact-4K, never-routed experts host-placed, both reference Triton kernels restored", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp1-placement-hctriton-qsafused-context4k-a305", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "New authority whose difference is at depth; quality profile byte-identical to the certified battery; lossless MTP1 within the lineage; LocalMaxxing 37.83 tok/s approved" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP0, exact-4K, never-routed experts host-placed, both reference Triton kernels restored", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp0-placement-hctriton-qsafused-context4k-a304", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "New authority whose difference is at depth; quality profile byte-identical to the certified battery; LocalMaxxing 33.80 tok/s approved" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 lossless MTP1, exact-4K, never-routed experts host-placed, Triton HC glue", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp1-placement-hctriton-context4k-a271", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "New deterministic authority reproduced on five servers; quality profile equal to the certified rows; lossless MTP1 within the lineage; LocalMaxxing 37.05 tok/s approved" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP0, exact-4K, never-routed experts host-placed, Triton HC glue", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp0-placement-hctriton-context4k-a269", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "New deterministic authority reproduced on three servers; quality profile equal to the certified rows; LocalMaxxing 32.90 tok/s approved" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 deterministic full-decode graph, exact-4K, never-routed experts host-placed (no speculation)", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp0-placement-context4k-a223", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Certified line: outputs bit-identical to the promoted authorities on two servers; 6/7 semantic with the inherited miss, 16/16 repeat, exact needle; LocalMaxxing 27.64 tok/s approved" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 lossless MTP1, exact-4K, never-routed experts host-placed", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp1-placement-context4k-a225", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Every output pin equal to the MTP0 line at exact 2K and 4K; certified battery; LocalMaxxing 31.93 tok/s approved" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 lossless MTP1 short context", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp1-fullgraphdet-short-a121", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Every output pin equal to the MTP0 line on two servers; 1.20x at short context, slower at depth" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP3 exact-4K", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp3-context4k-a1", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Exact-4K target-matched Grade-C screen; direct-answer target quality 6/7 semantic, with target-only official thinking separately 25/25" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP4 short screen", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp4-a1", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Configured-512 target-matched Grade-C screen; exact-4K MTP4 is separately quarantined" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP0 short screen", "measurement_id": "qwen38-flash-next-fp8-tp4-attempt19", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Configured-512 Grade-C target screen; direct-answer target quality 6/7 semantic, with target-only official thinking separately 25/25" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP1 short screen", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp1-a3", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Configured-512 target-matched Grade-C screen" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP2 short screen", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp2-a1", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Configured-512 target-matched Grade-C variable screen" }, { "role": "support", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP3 short screen", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp3-a4", "metric": "decode_tok_s", "sample_index": 0, "quality_label": "Configured-512 target-matched Grade-C variable screen" } ], "architecture": { "class": "Qwen4Exp hybrid Gated DeltaNet/QSA sparse MoE", "model_type": "qwen4_exp", "design": "125B total, about 6B active, plus 51B n-gram embedding and 4B MTP", "layers": 48, "hidden_size": 2560, "experts": 512, "experts_used": 10, "native_context_tokens": 262144, "evidence": "results/qwen38-flash-next-fp8-b70/README.md" }, "dimensions": { "weight_revision": [ "qwen38-flash-next" ], "weight_quantization": [ "FP8 block-128" ], "runtime": [ "vLLM XPU 658965050 + kernels 2f829747", "vLLM XPU 1372c62d + staged kernels 2f829747", "vLLM XPU 2169dbfe (overlay on 1372c62d) + staged kernels 2f829747", "vLLM XPU 1b2a17c1 (overlay on 1372c62d) + staged kernels 2f829747" ], "tp": [ 1, 2, 4 ], "ep": [ 1, 2, 4 ], "mtp": [ 0, 1, 2, 3, 4 ], "active_context_tokens": [ 0, 1024, 2048, 4096, 8192, 16384, 24576, 32768 ], "configured_max_context_tokens": [ 512, 1536, 3072, 4352, 8448, 262144 ], "graph_mode": [ "off", "PIECEWISE", "FULL_DECODE_ONLY" ], "kv": [ "auto" ], "modality": [ "text", "vision" ] }, "weight_revisions": [ { "id": "qwen38-flash-next", "label": "Qwen3.8 Flash-Next", "role": "base post-trained weights", "repository": "Qwen/Qwen3.8-Flash-Next", "revision_status": "The retained FP8 export identifies this base repository, but the exact parent base commit is not locally pinned.", "quantized_artifacts": [ { "id": "qwen38-flash-next-fp8-bcd9f01", "label": "Official fine-grained block-FP8 export", "quantization": "FP8 block-128", "quantization_origin": "export", "repository": "Qwen/Qwen3.8-Flash-Next-FP8", "revision": "bcd9f01ddc9cff2316eb84281bebcd5b058bddce", "evidence": "model-intake/post-download-validation-20260826.md" } ] } ], "model_variants": [], "transfer_scope": { "status": "The FP8 repository is a quantized child of Qwen3.8 Flash-Next, not a separate model. Measurements remain artifact- and runtime-specific.", "transfers": [ "the Qwen4Exp architecture integration and model-specific source patches across exact weight formats after independent qualification", "the artifact verification, cache-zero quality, and coverage workflow" ], "does_not_transfer": [ "attempt-19 performance or output behavior to unquantized weights, another export, or another runtime", "the exact-4K text screen to native 262K context or vision serving", "MTP0 evidence to any speculative depth" ], "evidence": "results/qwen38-flash-next-fp8-b70/README.md" }, "model_signals": { "b70_fit": { "band": "four-card screened", "scope": "TP4 plus EP4 with selective host placement on four 32-GiB B70 cards", "basis": "The exact FP8 export became healthy and served real text requests on all four cards.", "reviewed_at": "2026-08-27" }, "quality_evidence": { "band": "official-target-quality-pass-research-deployment", "scope": "The deterministic direct-answer target profile is 6/7 semantically, and the separate official-thinking MTP0 target profile passed 25/25. A current-runtime MTP0 replay independently passed short quality, 16/16 repeats, an exact cache-zero 4K needle, two repeat-identical exact-4K formal rows, and card-clean teardown. The retained legacy MTP0 profile has screened 1K/2K/4K and formal 8K gates; MTP1-4 have separate configured-512 and exact-4K evidence with MTP3 preferred at 4K. Active-context quarantines remain explicit: MTP1, MTP2, and MTP4 completed 8K but differ from the cross-runtime MTP0 authority, while MTP3 reached its fixed 900-second 8K bound without a completed receipt. The current-runtime 16K semantic program also failed same-server and fresh-server stability. No quarantined arm receives speed, quality, or deployment credit. Graph, vision, other topologies, clean-host qualification, and production deployment remain open.", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-attempt19-production-qualification.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-1536-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-3072-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-3072-context-repeat-v2-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-4352-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-8448-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-512-attempt3-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-1536-context-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-3072-context-attempt2-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-8448-context-attempt1-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-512-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-8448-context-attempt2-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-512-attempt4-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-8448-context-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-512-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-1536-context-attempt1-teardown-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-3072-context-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt1-pre-request-stop.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt2-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-4352-headroom32-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-4352-headroom32-attempt2-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-4352-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-official-quality-attempt2-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-official-quality-attempt2-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-4352-attempt1-mixed-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-4352-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-runtime-anchor-attempt4-result.json" ] }, "popularity": { "state": "dated-repository-snapshot", "captured_at": "2026-08-28T12:38:03Z", "repository": "Qwen/Qwen3.8-Flash-Next-FP8", "repository_revision": "970c569adaca6b35532111fd6b27351b2baefe50", "artifact_revision": "bcd9f01ddc9cff2316eb84281bebcd5b058bddce", "created_at": "2026-08-24T08:25:23.000Z", "last_modified": "2026-08-27T05:04:18.000Z", "downloads": 2219, "likes": 139, "scope": "Official FP8 repository-level discovery signal captured four days after repository creation; not model quality, artifact correctness, deployment maturity, or a revision-specific count. The only later official commit after the retained artifact is titled Update README.md.", "evidence": "data/neural-download-popularity/2026-08-28-qwen38-flash-next-fp8-huggingface.json" } }, "run_measurements": [ { "id": "qwen38-flash-next-fp8-tp4-attempt19", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 0, "configured_max_context_tokens": 512, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-production-source-screen-v1", "measurement_class": "research-only HTTP speed screen", "promotion_status": "not promotion eligible; short quality qualification failed", "quality_scope": "5/7 strict cases in both batteries; 15/16 repeat stability; substantive reasoning miss; no long-context, vision, or clean-host qualification", "workload": "One instrumentation-free server; p146/o256/c1; three sequential repetitive-prompt requests; full-output rate after first text, not the conventional 99-interval final gate", "metrics": { "decode_tok_s": [ 5.221849709057954 ], "ttft_ms": [ 5376.467669993872 ] }, "raw_observations": { "decode_tok_s": [ 5.142647219137399, 5.221849709057954, 5.289933931346957 ], "ttft_ms": [ 7752.22969299648, 5376.467669993872, 4535.219874989707 ], "aggregation": "median of three sequential same-server observations; all raw values remain in the evidence receipt" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 5.221849709057954, "label": "median research screen; short quality failed" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-attempt19-production-qualification.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-current-a4", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 0, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-current-anchor-a4-v1", "measurement_class": "research-only current-runtime HTTP speed and quality screen", "promotion_status": "not promotion eligible; short workload is not the fixed realistic final suite and fresh-server plus clean-host replay remain open", "quality_scope": "Current runtime passed 6/7 semantic cases with only the known code-expression miss, 16/16 same-output repeats, the exact cache-zero 4K needle, and card-clean teardown; this is same-boot Grade-C evidence", "workload": "Established p146/o256/c1 after-first-text screen; row one followed one conditioning request in its invocation and rows two and three had no warmup; the harness does not retain per-row cache detail or finish reason", "metrics": { "decode_tok_s": [ 5.223788770075911 ] }, "raw_observations": { "decode_tok_s": [ 5.315577823568996, 5.223788770075911, 5.2194047220586395 ], "usage_each": { "prompt_tokens": 146, "completion_tokens": 256, "total_tokens": 402 }, "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "aggregation": "median of three established after-first-text observations; raw values and scope boundaries remain in the tracked receipt" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 5.223788770075911, "label": "current-runtime short Grade-C research screen; not a final-suite headline" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-runtime-anchor-attempt4-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-current-context4k-a4", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-current-context4k-a4-v1", "measurement_class": "research-only repeated exact-depth HTTP screen", "promotion_status": "not promotion eligible; two repeats are same-boot and clean-host plus fresh-server replay remain open", "quality_scope": "Both exact-4K rows passed exact 4096/128/4224 usage, cache-zero, length-stop, token-count, fixture, and output-repeat gates after the current-runtime 6/7 semantic and 16/16 repeat battery", "workload": "Two exact p4096/o128 requests; conventional decode across the 99 inter-token intervals between generated-token events 1 and 100 after TTFT", "metrics": { "decode_tok_s": [ 4.7578181021380175 ], "ttft_ms": [ 147468.33551250165 ] }, "raw_observations": { "decode_tok_s": [ 4.720311369546584, 4.795324834729452 ], "ttft_ms": [ 149329.680180992, 145606.9908440113 ], "cached_tokens": [ 0, 0 ], "usage_each": { "prompt_tokens": 4096, "completion_tokens": 128, "total_tokens": 4224 }, "output_token_ids_sha256": "1d833e5f463366223a669aa15495840d1337b173e675a9ea04f00a5ae339d5cc", "aggregation": "median of two same-boot exact-depth observations; both raw rows remain hash-bound in the tracked receipt" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 4.7578181021380175, "label": "current-runtime exact-4K conventional Grade-C research screen" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-runtime-anchor-attempt4-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-a3", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 0, "configured_max_context_tokens": 512, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-matched-screen-v1", "measurement_class": "research-only target-matched MTP HTTP speed screen", "promotion_status": "not promotion eligible; inherited short quality remains 5/7", "quality_scope": "All 26 MTP0 baseline comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; small needle passed; all 24 quality requests cache-zero", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "metrics": { "decode_tok_s": [ 9.37225436776222 ], "ttft_ms": [ 9348.705686017638 ] }, "raw_observations": { "decode_tok_s": [ 9.773840621000225, 9.37225436776222, 8.107468408397532 ], "ttft_ms": [ 10140.820227010408, 8308.315517002484, 9348.705686017638 ], "mtp0_reference_median_tok_s": 5.221849709057954, "uplift_percent": 79.481503489173, "cumulative_draft_tokens": 505, "cumulative_accepted_tokens": 503, "aggregation": "median of three matched same-server observations; all raw values remain in the evidence receipt" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 9.37225436776222, "label": "median matched MTP1 research screen; inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-512-attempt3-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp2-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 2, "active_context_tokens": 0, "configured_max_context_tokens": 512, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp2-matched-screen-v1", "measurement_class": "research-only target-matched MTP HTTP speed screen", "promotion_status": "not promotion eligible; inherited short quality remains 5/7 and speed stability is unqualified", "quality_scope": "All 26 MTP0 baseline comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; small 317-token needle passed; all 24 quality requests complete and cache-zero", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup per invocation then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "metrics": { "decode_tok_s": [ 11.895061402541456 ], "wall_output_tok_s": [ 7.804965165044581 ], "ttft_ms": [ 11278.097242000513 ] }, "raw_observations": { "decode_tok_s": [ 13.586500711704096, 10.064084892145722, 11.895061402541456 ], "wall_output_tok_s": [ 8.57494723556275, 6.4119173926046695, 7.804965165044581 ], "ttft_ms": [ 11012.178995995782, 14488.66739301593, 11278.097242000513 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "row_span_percent_of_median": 29.61242233525408, "cumulative_draft_tokens": 770, "cumulative_accepted_tokens": 770, "aggregation": "median of three same-server observations; rows span 29.61% of the median, so cross-depth comparisons are descriptive rather than stable same-window causal estimates" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 11.895061402541456, "label": "median matched MTP2 research screen; variable rows and inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-512-attempt1-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp3-a4", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 3, "active_context_tokens": 0, "configured_max_context_tokens": 512, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp3-matched-screen-v1", "measurement_class": "research-only target-matched MTP HTTP speed screen", "promotion_status": "not promotion eligible; inherited short quality remains 5/7 and speed stability is unqualified", "quality_scope": "All 26 MTP0 baseline comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; small 317-token needle passed; all 24 quality requests complete and cache-zero; no 4K MTP3 claim", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup per invocation then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "metrics": { "decode_tok_s": [ 14.88878979448863 ], "ttft_ms": [ 11817.638449982041 ] }, "raw_observations": { "decode_tok_s": [ 17.47332085152878, 14.88878979448863, 12.538688913202236 ], "wall_output_tok_s": [ 9.671857481052726, 9.011438903429243, 7.592681229019125 ], "ttft_ms": [ 11817.638449982041, 11214.193463994889, 13299.87188798259 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "mtp0_reference_median_tok_s": 5.221849709057954, "mtp0_uplift_percent": 185.12482403815937, "mtp1_reference_median_tok_s": 9.37225436776222, "mtp1_uplift_percent": 58.86028281201647, "cumulative_draft_tokens": 768, "cumulative_accepted_tokens": 768, "aggregation": "median of three same-server observations; the rows declined monotonically and span 33.14% of the median, so cross-run uplift is descriptive rather than a stable same-window causal estimate" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 14.88878979448863, "label": "median matched MTP3 research screen; variable rows and inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-512-attempt4-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp4-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 4, "active_context_tokens": 0, "configured_max_context_tokens": 512, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp4-matched-screen-v1", "measurement_class": "research-only target-matched MTP HTTP speed screen", "promotion_status": "not promotion eligible; inherited short quality remains 5/7 and deployment stability is unqualified", "quality_scope": "All 26 MTP0 baseline comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; small 317-token needle passed; all 24 quality requests complete and cache-zero", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup per invocation then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "metrics": { "decode_tok_s": [ 20.72717637199404 ], "wall_output_tok_s": [ 11.560326762555018 ], "ttft_ms": [ 10023.315081984038 ] }, "raw_observations": { "decode_tok_s": [ 21.119694109018415, 18.57624860519203, 20.72717637199404 ], "wall_output_tok_s": [ 11.560326762555018, 10.360395438896376, 12.276512942462682 ], "ttft_ms": [ 10023.315081984038, 10928.44290501671, 8501.892345986562 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "row_span_percent_of_median": 12.271066054433797, "mtp3_reference_median_tok_s": 14.88878979448863, "mtp3_descriptive_uplift_percent": 39.213305165115564, "cumulative_draft_tokens": 1716, "cumulative_accepted_tokens": 1716, "aggregation": "median of three corrected same-server observations; rows span 12.27% of the median and the receipt preserves the preceding filename-loop diagnostic artifact" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 20.72717637199404, "label": "median matched MTP4 research screen; inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-512-attempt1-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-context4k-headroom32-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-context4k-headroom32-screen-v1", "measurement_class": "research-only exact-4K HTTP service screen with 32-block cache headroom and separate formal row", "promotion_status": "not promotion eligible; inherited short quality remains 5/7 and clean-boot/deployment stability are unqualified", "quality_scope": "All 26 sealed MTP0 4K comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; exact 4096-token needle and formal depth gate passed; all 24 quality requests cache-zero; three deployment-shaped rows passed", "workload": "Three separately salted exact p4096/o256/c1 requests with no harness-added warmups; full-output rate after first text; formal p4096/o128 99-interval row retained separately", "metrics": { "decode_tok_s": [ 8.904420575355882 ], "ttft_ms": [ 232079.23328102333 ], "wall_output_tok_s": [ 0.9810504777261468 ] }, "raw_observations": { "decode_tok_s": [ 8.904420575355882, 8.868704696563922, 9.581812273689444 ], "wall_output_tok_s": [ 0.8762526385395759, 0.9810504777261468, 1.1657254437993072 ], "ttft_ms": [ 263403.42078500544, 232079.23328102333, 192888.45706998836 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "fixed_allocation_blocks": 32, "reported_cache_capacity_tokens": 9284, "formal_99_interval_tok_s": 3.4714510192622985, "formal_ttft_ms": 317104.6654289821, "decode_row_span_percent_of_median": 8.00847, "wall_output_row_span_percent_of_median": 29.506412955495, "ttft_row_span_percent_of_median": 30.384004082619, "cumulative_draft_tokens": 539, "cumulative_accepted_tokens": 528, "aggregation": "median of three exact-4K service observations; the 32-block allocation is a recipe parameter, not proof that cache headroom caused the MTP2/MTP4 stalls to disappear" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 8.904420575355882, "label": "median exact-4K MTP1 headroom32 Grade-C screen; 232.1-second TTFT and inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-4352-headroom32-attempt1-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp2-context4k-headroom32-a2", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 2, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp2-context4k-headroom32-screen-v1", "measurement_class": "research-only exact-4K HTTP service screen with 32-block cache headroom and separate formal row", "promotion_status": "not promotion eligible; inherited short quality remains 5/7 and clean-boot/deployment stability are unqualified", "quality_scope": "All 26 sealed MTP0 4K comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; exact 4096-token needle and formal depth gate passed; all 24 quality requests cache-zero; three deployment-shaped rows passed", "workload": "Three separately salted exact p4096/o256/c1 requests with no harness-added warmups; full-output rate after first text; formal p4096/o128 99-interval row retained separately", "metrics": { "decode_tok_s": [ 9.89315479235244 ], "ttft_ms": [ 263279.2244020093 ], "wall_output_tok_s": [ 0.891381690144734 ] }, "raw_observations": { "decode_tok_s": [ 9.89315479235244, 12.07804962791263, 9.217263500488167 ], "wall_output_tok_s": [ 0.8543969736295236, 0.8999042813260602, 0.891381690144734 ], "ttft_ms": [ 273750.0517080189, 263279.2244020093, 259420.6210770062 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "fixed_allocation_blocks": 32, "reported_cache_capacity_tokens": 7329, "maximum_observed_cache_usage_percent": 61.3, "formal_99_interval_tok_s": 3.4792396609382594, "formal_ttft_ms": 369141.1535299849, "decode_row_span_percent_of_median": 28.916823677275268, "wall_output_row_span_percent_of_median": 5.105254931717019, "ttft_row_span_percent_of_median": 5.442674279962402, "cumulative_draft_tokens": 748, "cumulative_accepted_tokens": 719, "aggregation": "median of three exact-4K service observations; compared with the failed 21-block arm, only the fixed cache allocation changed. The result supports this recipe but does not prove 32 blocks is minimal or explain other MTP depths." }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 9.89315479235244, "label": "median exact-4K MTP2 headroom32 Grade-C screen; 263.3-second TTFT and inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-4352-headroom32-attempt2-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp3-context4k-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 3, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp3-context4k-screen-v1", "measurement_class": "research-only exact-4K HTTP service screen with separate formal row", "promotion_status": "not promotion eligible; inherited short quality remains 5/7 and TTFT/deployment stability are unqualified", "quality_scope": "All 26 sealed MTP0 4K comparisons matched; inherited strict score 5/7; fixed-set repeats 16/16 with one hash; exact 4096-token needle and formal depth gate passed; all 24 quality requests cache-zero", "workload": "Three separately salted exact p4096/o256/c1 requests with no harness-added warmups; full-output rate after first text; formal p4096/o128 99-interval row retained separately", "metrics": { "decode_tok_s": [ 15.50156510641242 ], "ttft_ms": [ 187899.1858829977 ], "wall_output_tok_s": [ 1.2462600034136797 ] }, "raw_observations": { "decode_tok_s": [ 16.57897611005858, 15.50156510641242, 14.615697889304094 ], "wall_output_tok_s": [ 1.2834112898631422, 1.2456291788090392, 1.2462600034136797 ], "ttft_ms": [ 184027.15963101946, 189004.16664901422, 187899.1858829977 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "formal_99_interval_tok_s": 4.669548248983529, "formal_ttft_ms": 266080.89533800376, "mtp0_legacy_reference_median_tok_s": 5.233664731906276, "mtp0_legacy_decode_uplift_percent": 196.18949436920138, "mtp0_legacy_ttft_delta_percent": 52.279150769899616, "mtp0_legacy_wall_delta_percent": -16.118243212788563, "cumulative_draft_tokens": 852, "cumulative_accepted_tokens": 799, "aggregation": "median of three exact-4K service observations; decode is reported with TTFT and wall throughput because long-prompt prefill dominates end-to-end latency" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 15.50156510641242, "label": "median exact-4K MTP3 research screen; 187.9-second TTFT and inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-4352-attempt1-result.json" }, { "id": "qwen38-flash-next-fp8-tp4-context1k-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 1024, "configured_max_context_tokens": 1536, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-context1k-screen-v1", "measurement_class": "research-only exact-depth HTTP screen", "promotion_status": "not promotion eligible; inherited short quality qualification failed", "quality_scope": "987-token exact needle and 16/16 repeats passed; 12-prompt realistic timing gate passed cache-zero; short battery remained 5/7", "workload": "One instrumentation-free server; three unique exact p1024/o256/c1 requests; full-output rate after first text; no harness-added warmups after the server completed prerequisite gates; separate 12-prompt realistic suite", "metrics": { "decode_tok_s": [ 5.13358756138473 ], "ttft_ms": [ 29043.11463900376 ] }, "raw_observations": { "decode_tok_s": [ 5.153794240763608, 5.13358756138473, 5.0512700513432325 ], "ttft_ms": [ 29686.471389984945, 25229.969660984352, 29043.11463900376 ], "realistic_99_interval_median_tok_s": 4.449168445347031, "aggregation": "median of three unique exact-1K observations; raw values and realistic outputs remain in the evidence receipt" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 5.13358756138473, "label": "median exact-1K research screen; short quality failed" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-1536-context-screen.json" }, { "id": "qwen38-flash-next-fp8-tp4-context2k-v2-a2", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 2048, "configured_max_context_tokens": 3072, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-context2k-repeat-v2-screen-v1", "measurement_class": "research-only exact-depth HTTP screen", "promotion_status": "not promotion eligible; inherited short quality qualification failed", "quality_scope": "Protocol-v2 fixed-set first-token sensitivity passed 32/32, full repeats passed 16/16 with one exact output, and the formal 2K cache-zero gate passed; short battery remained 5/7", "workload": "One instrumentation-free server; three unique exact p2048/o256/c1 requests; full-output rate after first text; no harness-added warmups after the server completed prerequisite gates; separate formal exact-p2048/o128 cache-zero gate", "metrics": { "decode_tok_s": [ 5.228429046201661 ], "ttft_ms": [ 81146.86629301286 ] }, "raw_observations": { "decode_tok_s": [ 5.034312884410624, 5.257401637348096, 5.228429046201661 ], "ttft_ms": [ 87998.60319800791, 81146.86629301286, 76276.06647202629 ], "formal_99_interval_tok_s": 3.8648778892817663, "formal_ttft_ms": 111659.75860299659, "aggregation": "median of three unique exact-2K legacy-comparable observations; formal cache-zero rate is retained separately in the evidence receipt" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 5.228429046201661, "label": "median exact-2K protocol-v2 research screen; short quality failed" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-3072-context-repeat-v2-screen.json" }, { "id": "qwen38-flash-next-fp8-tp4-context4k-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-context4k-screen-v1", "measurement_class": "research-only formal exact-depth HTTP screen with secondary legacy comparison", "promotion_status": "not promotion eligible; inherited short quality qualification failed", "quality_scope": "All seven short outputs matched the 2K baseline; fixed-set repeats passed 16/16 with one exact output; the exact-4K needle and formal cache-zero gate passed; short battery remained 5/7", "workload": "Primary formal cache-zero exact-p4096/o128 100-event/99-interval row; secondary three-prompt exact-p4096/o256 legacy after-first-text comparison on the same instrumentation-free server", "metrics": { "decode_tok_s": [ 4.4560264746397324 ], "ttft_ms": [ 217909.69186701113 ] }, "raw_observations": { "legacy_after_first_text_decode_tok_s": [ 5.298983874891171, 5.233664731906276, 5.161604624443941 ], "legacy_after_first_text_ttft_ms": [ 132052.37051899894, 123391.27512401319, 117952.53843598766 ], "legacy_after_first_text_median_tok_s": 5.233664731906276, "legacy_after_first_text_median_ttft_ms": 123391.27512401319, "formal_99_interval_tok_s": 4.4560264746397324, "formal_ttft_ms": 217909.69186701113, "aggregation": "primary metrics use the formal cache-zero row; legacy after-first-text comparison is retained separately for the context curve" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 4.4560264746397324, "label": "formal exact-4K cache-zero 99-interval research screen; short quality failed" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-4352-context-screen.json" }, { "id": "qwen38-flash-next-fp8-tp4-context4k-service-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-context4k-service-v1", "measurement_class": "research-only exact-4K HTTP service comparison", "promotion_status": "not promotion eligible; inherited short quality remains 5/7", "quality_scope": "Same exact-4K MTP0 gate as the formal measurement; this separate workload record retains only its comparable p4096/o256 service rows", "workload": "Three separately salted exact p4096/o256/c1 requests with no harness-added warmups; full-output rate after first text; formal p4096/o128 row retained separately", "metrics": { "decode_tok_s": [ 5.233664731906276 ], "ttft_ms": [ 123391.27512401319 ], "wall_output_tok_s": [ 1.485734265884717 ] }, "raw_observations": { "decode_tok_s": [ 5.298983874891171, 5.233664731906276, 5.161604624443941 ], "wall_output_tok_s": [ 1.4193557575079974, 1.485734265884717, 1.5279065125481308 ], "ttft_ms": [ 132052.37051899894, 123391.27512401319, 117952.53843598766 ], "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "aggregation": "median of three exact-4K service observations; retained separately from the formal p4096/o128 measurement" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 5.233664731906276, "label": "median exact-4K MTP0 service comparison; inherited short quality 5/7" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-4352-context-screen.json" }, { "id": "qwen38-flash-next-fp8-tp4-context8k-a1", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 8192, "configured_max_context_tokens": 8448, "graph_mode": "off", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-context8k-screen-v1", "measurement_class": "research-only formal exact-depth HTTP screen with incomplete secondary comparison", "promotion_status": "not promotion eligible; inherited short quality and repeated-comparison stability gates failed", "quality_scope": "All seven short outputs matched the 4K baseline; fixed-set repeats passed 16/16 with one exact output; the exact-8K needle and formal cache-zero gate passed; short battery remained 5/7; the runtime stopped during the third legacy-comparison request", "workload": "Primary formal cache-zero exact-p8192/o128 100-event/99-interval row; two valid exact-p8192/o256 legacy observations retained without a median because the required third row did not complete", "metrics": { "decode_tok_s": [ 3.97972923995132 ], "ttft_ms": [ 386534.3322980043 ] }, "raw_observations": { "legacy_after_first_text_decode_tok_s_valid_rows": [ 5.170404147374639, 5.182352525810221 ], "legacy_after_first_text_ttft_ms_valid_rows": [ 243262.31589599047, 216963.4954840003 ], "legacy_median_authorized": false, "legacy_third_row": "invalid: no text or usage before runtime shutdown", "formal_99_interval_tok_s": 3.97972923995132, "formal_ttft_ms": 386534.3322980043, "aggregation": "primary metrics use the completed formal cache-zero row; two legacy observations remain raw only and no 8K legacy curve point is authorized" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 3.97972923995132, "label": "formal exact-8K cache-zero 99-interval research screen; repeated comparison incomplete and short quality failed" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-8448-context-screen.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-fullgraphdet-short-a78", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 2169dbfe (overlay on 1372c62d) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 0, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-fullgraphdet-ctx0-v1", "measurement_class": "promoted frozen-client short p146/o256 rows on the deterministic full-decode-graph line", "promotion_status": "deterministic line promoted as the lab TP4 record on 2026-09-03 (results packet section 'Deterministic full-decode-graph line'); LocalMaxxing cmtn32b2w000tmm01t7j2wlpn approved 2026-09-04 on the fixed realistic suite (14.433684 tok/s class-balanced median)", "quality_scope": "6/7 semantic (sole miss code_execution=30), 16/16 repeat with one hash, exact cache-zero 2K needle; outputs identical across five independently started servers (A70-A73, A78); first-step logits and 128-token continuations bit-identical across repeats at depths 8-4096", "workload": "three separately salted p146/o256/c1 requests (row 1 after one conditioning request); rate after first text; two servers A73 and A78", "metrics": { "decode_tok_s": [ 22.660696 ] }, "raw_observations": { "decode_tok_s": [ 22.966002, 23.898996, 22.256402, 22.35539, 23.350884, 22.321053 ], "aggregation": "median of the two attempt medians (A73 22.966002, A78 22.355390); six rows 22.26-23.90", "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 22.660696, "label": "deterministic full-decode-graph line, no speculation; center of two servers" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260903-tp4-mtp0-a78-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-fullgraphdet-context2k-a78", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 2169dbfe (overlay on 1372c62d) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 2048, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-fullgraphdet-ctx2048-v1", "measurement_class": "promoted frozen-client exact-2K p2048/o128 rows (conventional 99 inter-token intervals)", "promotion_status": "deterministic line promoted as the lab TP4 record on 2026-09-03 (results packet section 'Deterministic full-decode-graph line'); LocalMaxxing cmtn32b2w000tmm01t7j2wlpn approved 2026-09-04 on the fixed realistic suite (14.433684 tok/s class-balanced median)", "quality_scope": "6/7 semantic (sole miss code_execution=30), 16/16 repeat with one hash, exact cache-zero 2K needle; outputs identical across five independently started servers (A70-A73, A78); first-step logits and 128-token continuations bit-identical across repeats at depths 8-4096", "workload": "two exact p2048/o128 requests per server, cache zero; 99-interval rate; TTFT recorded", "metrics": { "decode_tok_s": [ 13.993164 ], "ttft_ms": [ 57167.8 ] }, "raw_observations": { "decode_tok_s": [ 13.514374, 14.909545, 13.44331, 14.471953 ], "ttft_ms": [ 58256.2, 56079.4 ], "aggregation": "median of four rows on two servers", "output_sha256": "afffd2110812762164862b6388f054bb56696ee57b07eadce411a702c40bc714" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 13.993164, "label": "deterministic line exact-2K median; hash afffd211 on seven servers" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260903-tp4-mtp0-a78-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-fullgraphdet-context4k-a78", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 2169dbfe (overlay on 1372c62d) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-fullgraphdet-ctx4096-v1", "measurement_class": "promoted frozen-client exact-4K p4096/o128 rows (conventional 99 inter-token intervals)", "promotion_status": "deterministic line promoted as the lab TP4 record on 2026-09-03 (results packet section 'Deterministic full-decode-graph line'); LocalMaxxing cmtn32b2w000tmm01t7j2wlpn approved 2026-09-04 on the fixed realistic suite (14.433684 tok/s class-balanced median)", "quality_scope": "6/7 semantic (sole miss code_execution=30), 16/16 repeat with one hash, exact cache-zero 2K needle; outputs identical across five independently started servers (A70-A73, A78); first-step logits and 128-token continuations bit-identical across repeats at depths 8-4096", "workload": "two exact p4096/o128 requests per server, cache zero; 99-interval rate; TTFT recorded", "metrics": { "decode_tok_s": [ 12.77677 ], "ttft_ms": [ 99436.9 ] }, "raw_observations": { "decode_tok_s": [ 12.728316, 12.825225, 13.498466, 12.241721 ], "ttft_ms": [ 98680.0, 89510.0, 102516.5, 96357.4 ], "aggregation": "median of four rows on two servers", "output_sha256": "c6193cc6c9a1553f56d7ce78faea9c8bfa628a67fcea229b1c99279a149f6639" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 12.77677, "label": "deterministic line exact-4K median, no speculation: 2.4-2.7x the native eager line" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260903-tp4-mtp0-a78-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-fullgraphdet-short-a121", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1b2a17c1 (overlay on 1372c62d) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 0, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-fullgraphdet-ctx0-v1", "measurement_class": "certified lossless frozen-client short p146/o256 rows with one speculative token", "promotion_status": "certified lossless on two servers 2026-09-03; short-context candidate (faster than the MTP0 line only at short context); MTP0 line remains the record at depth", "quality_scope": "Same gates; every output pin equals the MTP0 deterministic line's (short 5f407446, exact-2K afffd211, exact-4K c6193cc6, repeat 3b0b3192) on two independently started servers (A120, A121); the two-row verification step traced bit-identical to the MTP0 step through all 48 layers on every rank (A112)", "workload": "same short protocol as the MTP0 line; three exact-verify selectors (serial GDN verifier rows, row-wise TP all-reduce, row-wise hyperconnection norm variance)", "metrics": { "decode_tok_s": [ 27.149928 ] }, "raw_observations": { "decode_tok_s": [ 22.185575, 28.389615, 26.72301, 22.024228, 31.381879, 27.576847 ], "aggregation": "median of the two attempt medians (A120 26.723010, A121 27.576847); six rows 22.02-31.38; the cold-first diagnostic battery A113 gave 31.20/34.73/31.31", "output_sha256": "5f40744644b98ddd58a0c202fe855af324c0b1c33e1a6275afd74c12488f89f0", "draft_acceptance": "735 of 785 draft tokens over the A113 battery" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 27.149928, "label": "lossless MTP1: 1.20x the MTP0 line at short context in the frozen client (1.38x in the cold battery)" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260903-tp4-mtp1-a121-frozen-client-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-fullgraphdet-context2k-a121", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1b2a17c1 (overlay on 1372c62d) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 2048, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-fullgraphdet-ctx2048-v1", "measurement_class": "certified lossless frozen-client exact-2K rows with one speculative token", "promotion_status": "certified lossless on two servers 2026-09-03; short-context candidate (faster than the MTP0 line only at short context); MTP0 line remains the record at depth", "quality_scope": "Same gates; every output pin equals the MTP0 deterministic line's (short 5f407446, exact-2K afffd211, exact-4K c6193cc6, repeat 3b0b3192) on two independently started servers (A120, A121); the two-row verification step traced bit-identical to the MTP0 step through all 48 layers on every rank (A112)", "workload": "two exact p2048/o128 requests per server; 99-interval rate", "metrics": { "decode_tok_s": [ 9.039214 ], "ttft_ms": [ 88014.0 ] }, "raw_observations": { "decode_tok_s": [ 8.924331, 8.839258, 9.17304, 9.154097 ], "ttft_ms": [ 94694.8, 83542.1, 95443.7, 78309.8 ], "aggregation": "median of four rows on two servers", "output_sha256": "afffd2110812762164862b6388f054bb56696ee57b07eadce411a702c40bc714" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 9.039214, "label": "lossless MTP1 at 2K is 0.65x the MTP0 line: the two-row graph replay costs 144-255 ms per step against 71 ms (depth-cost note)" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260903-tp4-mtp1-a121-frozen-client-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-fullgraphdet-context4k-a121", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1b2a17c1 (overlay on 1372c62d) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-fullgraphdet-ctx4096-v1", "measurement_class": "certified lossless frozen-client exact-4K rows with one speculative token", "promotion_status": "certified lossless on two servers 2026-09-03; short-context candidate (faster than the MTP0 line only at short context); MTP0 line remains the record at depth", "quality_scope": "Same gates; every output pin equals the MTP0 deterministic line's (short 5f407446, exact-2K afffd211, exact-4K c6193cc6, repeat 3b0b3192) on two independently started servers (A120, A121); the two-row verification step traced bit-identical to the MTP0 step through all 48 layers on every rank (A112)", "workload": "two exact p4096/o128 requests per server; 99-interval rate", "metrics": { "decode_tok_s": [ 7.724194 ], "ttft_ms": [ 146706.0 ] }, "raw_observations": { "decode_tok_s": [ 7.140176, 7.433851, 8.014537, 8.05301 ], "ttft_ms": [ 149700.1, 145169.2, 151646.3, 140113.5 ], "aggregation": "median of four rows on two servers", "output_sha256": "c6193cc6c9a1553f56d7ce78faea9c8bfa628a67fcea229b1c99279a149f6639" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 7.724194, "label": "lossless MTP1 at 4K is 0.60x the MTP0 line" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260903-tp4-mtp1-a121-frozen-client-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-placement-context4k-a223", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU cb59004b (placement overlay on 2169dbfe) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-placement-ctx4096-v1", "measurement_class": "certified frozen-client exact-4K p4096/o128 rows (conventional 99 inter-token intervals) with never-routed experts host-placed", "promotion_status": "certified on two servers 2026-09-06 (A223, A224); LocalMaxxing cmtq4elns03ivn701hgvpo053 approved 2026-09-06 on the fixed realistic suite (27.640875 tok/s class-balanced median); supersedes the 25.62 headroom row as the fastest no-speculation line", "quality_scope": "6/7 semantic (sole miss code_execution=30), 16/16 repeat with one hash, exact cache-zero 2K needle; exact-2K afffd211 and exact-4K c6193cc6 authorities reproduced on both servers; tuned W13/N32 map selected on every layer (verifier receipt)", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table", "metrics": { "decode_tok_s": [ 27.397903 ] }, "raw_observations": { "decode_tok_s": [ 27.398396, 27.426248, 27.397409, 27.359961 ], "aggregation": "median of four rows on two servers (A223 r1/r2, A224 r1/r2)", "output_sha256": "c6193cc6c9a1553f56d7ce78faea9c8bfa628a67fcea229b1c99279a149f6639" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 27.397903, "label": "never-routed experts host-placed, hot experts resident: 1.08x the headroom line, bit-identical" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp0-a223-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-placement-context4k-a225", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 005dc578 (placement overlay on the lossless MTP1 head 1b2a17c1) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-placement-ctx4096-v1", "measurement_class": "certified lossless frozen-client exact-4K p4096/o128 rows with one speculative token and never-routed experts host-placed", "promotion_status": "certified lossless 2026-09-06 (A225: every output pin equal to the MTP0 line); LocalMaxxing cmtq59cy503jvn701kgvg62zt approved 2026-09-06 on the fixed realistic suite (31.929484 tok/s class-balanced median), the fastest quality-preserving Flash-Next row; record gate replayed end-to-end 2026-09-06 (attempt 229: 12/12 outputs identical, 32.18 tok/s)", "quality_scope": "6/7 semantic (sole miss code_execution=30), 16/16 repeat with one hash, exact needle; exact-2K afffd211 and exact-4K c6193cc6 authorities reproduced with MTP1 (lossless)", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table", "metrics": { "decode_tok_s": [ 32.489874 ] }, "raw_observations": { "decode_tok_s": [ 32.479529, 32.50022 ], "aggregation": "median of two rows on one server (A225 r1/r2)", "output_sha256": "c6193cc6c9a1553f56d7ce78faea9c8bfa628a67fcea229b1c99279a149f6639" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 32.489874, "label": "lossless MTP1 with never-routed experts host-placed: 1.19x the MTP0 placement line at exact 4K" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a225-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-placement-hctriton-context4k-a269", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 8d7d6fd8 (Triton-HC overlay on the MTP0 placement head cb59004b) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-placement-hctriton-ctx4096-v1", "measurement_class": "certified frozen-client exact-4K p4096/o128 rows (conventional 99 inter-token intervals) with never-routed experts host-placed", "promotion_status": "certified 2026-09-07 as a new Triton-HC authority (A269: every pin deterministic and equal to the A266/A267 screens; outputs not bit-identical to the torch-fallback rows); LocalMaxxing cmtqqyulr006fpa01erfuiri8 approved 2026-09-07 on the fixed realistic suite (32.898806 tok/s class-balanced median)", "quality_scope": "6/7 semantic (sole miss code_execution), byte-identical outputs to the certified placement battery on all seven cases, 16/16 repeat with one hash, exact needle; exact-2K 86b5b6c7 and exact-4K b89822ce (new authority) reproduced on three servers", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table; the Triton hyper-connection glue kernels run on XPU (VLLM_XPU_HC_TRITON=1)", "metrics": { "decode_tok_s": [ 32.595981 ] }, "raw_observations": { "decode_tok_s": [ 32.576113, 32.615849 ], "aggregation": "median of two rows on one server (A269 r1/r2)", "output_sha256": "b89822ce3b8a2719e0824d0e49222d535c8bd9fd86bc58feebd3cddaa9477b7c" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 32.595981, "label": "Triton HC glue on the MTP0 placement line: 1.19x the placement row at exact 4K, new output authority" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp0-a269-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-placement-hctriton-context4k-a271", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 62219122 (Triton-HC overlay on the lossless MTP1 placement head 005dc578) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-placement-hctriton-ctx4096-v1", "measurement_class": "certified (new Triton-HC authority) frozen-client exact-4K p4096/o128 rows with one speculative token and never-routed experts host-placed", "promotion_status": "certified 2026-09-07 (A271: every pin equal to the Triton-HC MTP0 line, lossless MTP1 within the new authority); LocalMaxxing cmtqspsy00092pa01kli5htlb approved 2026-09-07 on the fixed realistic suite (37.045844 tok/s class-balanced median), the fastest Flash-Next row; outputs not bit-identical to the torch-fallback rows; record gate replayed end-to-end 2026-09-07 (attempt 273: 12/12 outputs identical, 37.43 tok/s)", "quality_scope": "6/7 semantic (sole miss code_execution), byte-identical outputs to the certified placement battery on all seven cases, 16/16 repeat with one hash, exact needle; exact-2K 86b5b6c7 and exact-4K b89822ce authorities reproduced with MTP1 (lossless within the lineage)", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table; the Triton hyper-connection glue kernels run on XPU (VLLM_XPU_HC_TRITON=1)", "metrics": { "decode_tok_s": [ 36.365674 ] }, "raw_observations": { "decode_tok_s": [ 36.368651, 36.362698 ], "aggregation": "median of two rows on one server (A271 r1/r2)", "output_sha256": "b89822ce3b8a2719e0824d0e49222d535c8bd9fd86bc58feebd3cddaa9477b7c" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 36.365674, "label": "lossless MTP1 with the Triton HC glue: the fastest certified Flash-Next row at exact 4K, new output authority" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp1-a271-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-placement-hctriton-qsafused-context4k-a304", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 2a372e86 (fused-QSA overlay on the Triton-HC MTP0 head) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-placement-hctriton-qsafused-ctx4096-v1", "measurement_class": "certified frozen-client exact-4K p4096/o128 rows (conventional 99 inter-token intervals) with never-routed experts host-placed", "promotion_status": "certified 2026-09-07 (A304: all gates passed on a second server); LocalMaxxing cmtrmp37v001bps01a7fi46nf approved 2026-09-07 on the fixed realistic suite (33.797067 tok/s class-balanced); new authority whose difference is at depth (exact-2K coincides with the certified stream, exact-4K does not)", "quality_scope": "6/7 semantic (sole miss code_execution), all seven exact-case outputs byte-identical to the certified battery, 16/16 repeat one hash, exact needle; exact-2K afffd211 (coincides with the certified stream) and exact-4K 1d833e5f (new)", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table; the Triton hyper-connection glue kernels run on XPU (VLLM_XPU_HC_TRITON=1); the hyper-connection glue and the QSA pre-indexer both run the model's reference Triton kernels", "metrics": { "decode_tok_s": [ 33.44 ] }, "raw_observations": { "decode_tok_s": [ 33.47, 33.41 ], "aggregation": "median of two rows on one server (A304 r1/r2)", "output_sha256": "1d833e5f463366223a669aa15495840d1337b173e675a9ea04f00a5ae339d5cc" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 33.44, "label": "both reference Triton kernels restored on the MTP0 line: 1.22x the placement row at exact 4K" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp0-a304-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-placement-hctriton-qsafused-context4k-a305", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 6d872457 (fused-QSA overlay on the Triton-HC MTP1 head) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 1, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp1-placement-hctriton-qsafused-ctx4096-v1", "measurement_class": "certified (new Triton-HC authority) frozen-client exact-4K p4096/o128 rows with one speculative token and never-routed experts host-placed", "promotion_status": "certified 2026-09-07 (A305: all gates passed; pins equal to this lineage's MTP0 line, so MTP1 stays lossless within it); LocalMaxxing cmtrmp3mj001fps01thcathd0 approved 2026-09-07 on the fixed realistic suite (37.825654 tok/s class-balanced), the fastest Flash-Next row; new authority whose difference is at depth; record gate replayed end-to-end 2026-09-07 (attempt 307: 12/12 outputs identical, 37.97 tok/s)", "quality_scope": "6/7 semantic (sole miss code_execution), all seven exact-case outputs byte-identical to the certified battery, 16/16 repeat one hash, exact needle; exact-2K afffd211 and exact-4K 1d833e5f reproduced with MTP1 (lossless within the lineage)", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table; the Triton hyper-connection glue kernels run on XPU (VLLM_XPU_HC_TRITON=1); the hyper-connection glue and the QSA pre-indexer both run the model's reference Triton kernels", "metrics": { "decode_tok_s": [ 39.3 ] }, "raw_observations": { "decode_tok_s": [ 39.3, 39.3 ], "aggregation": "median of two rows on one server (A305 r1/r2)", "output_sha256": "1d833e5f463366223a669aa15495840d1337b173e675a9ea04f00a5ae339d5cc" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 39.3, "label": "lossless MTP1 with both reference kernels restored: the fastest certified Flash-Next row at exact 4K" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp1-a305-fresh-repeat-deterministic-summary.json" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-placement-hctriton-qsafused-w13n64-context4k-a326", "state": "lab-measured", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "variant": "official FP8 block-128", "quantization": "FP8 block-128", "runtime": "vLLM XPU 2a372e86 (fused-QSA overlay on the Triton-HC MTP0 head) + staged kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 4096, "configured_max_context_tokens": 4352, "graph_mode": "FULL_DECODE_ONLY", "kv": "auto", "modality": "text" }, "profile_id": "flash-next-tp4-mtp0-placement-hctriton-qsafused-w13n64-ctx4096-v1", "measurement_class": "certified fixed cold realistic-suite median (class-balanced, 99 inter-token intervals after TTFT) with never-routed experts host-placed and the W13 phase tile neutralised to the base 64", "promotion_status": "certified 2026-09-08 (A325 34.510128 and A326 34.495292 on two cold servers, every row above every row of the superseded A301 suite; exact-2K and exact-4K hashes unmoved across A321/A322/A323 on three servers). One line of the tuned M1 map: W1_CONFIG.BLOCK_SIZE_N 32 -> 64, making the W13 phase delta equal to the base tile and therefore inert; no source change, certified head 2a372e86 unmodified. LocalMaxxing cmts8zca50032ps01e0ddqm18 approved 2026-09-08. Supersedes the 33.797067 measurement, which is retained.", "quality_scope": "6/7 semantic (sole miss code_execution), all seven exact-case outputs byte-identical to the certified battery, 16/16 repeat one hash, exact needle; exact-2K afffd211 (coincides with the certified stream) and exact-4K 1d833e5f (new)", "workload": "two exact p4096/o128 requests per server on two servers, cache zero; 99-interval rate; PLE and embeddings host-offloaded over UVA (12.22 GiB/rank), every hot routed expert resident, 3.5 GiB/rank of never-routed experts in pinned host memory behind a per-expert offset table; the Triton hyper-connection glue kernels run on XPU (VLLM_XPU_HC_TRITON=1); the hyper-connection glue and the QSA pre-indexer both run the model's reference Triton kernels", "metrics": { "decode_tok_s": [ 33.44 ] }, "raw_observations": { "decode_tok_s": [ 33.47, 33.41 ], "aggregation": "median of two rows on one server (A304 r1/r2)", "output_sha256": "1d833e5f463366223a669aa15495840d1337b173e675a9ea04f00a5ae339d5cc" }, "sample_annotations": [ { "metric": "decode_tok_s", "index": 0, "value": 33.44, "label": "both reference Triton kernels restored on the MTP0 line: 1.22x the placement row at exact 4K" } ], "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp0-a304-fresh-repeat-deterministic-summary.json" } ], "series_measurements": [ { "id": "qwen38-flash-next-fp8-tp4-context-screen", "state": "lab-screened", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "variant": "official FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "config": { "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "workload": "HTTP p146/o256 control median at x=0 plus three unique exact-p1024/o256, exact-p2048/o256, and exact-p4096/o256 medians; after-first-text legacy-comparable rates; research-only quality boundary", "axis": "active_context_tokens", "points": [ { "x": 0, "decode_tok_s": 5.221849709057954, "samples": 3 }, { "x": 1024, "decode_tok_s": 5.13358756138473, "ttft_ms": 29043.11463900376, "samples": 3 }, { "x": 2048, "decode_tok_s": 5.228429046201661, "ttft_ms": 81146.86629301286, "samples": 3 }, { "x": 4096, "decode_tok_s": 5.233664731906276, "ttft_ms": 123391.27512401319, "samples": 3 } ], "quality": "All plotted points remain research-only because the same substantive short-quality miss reproduced. The 1K point passed its needle, repeats, and cache-zero realistic suite; the 2K and 4K points passed exact needles, fixed-set repeats, and separate formal cache-zero exact-depth gates. The 8K formal gate also passed, but it is excluded from this legacy-comparable series because the required third comparison row did not complete.", "caveat": "The x=0 control used a 146-token ordinary prompt at configured max 512; x=1024 used exact p1024 at configured max 1536; x=2048 used exact p2048 at configured max 3072; x=4096 used exact p4096 at configured max 4352. Formal 99-interval results are separately labeled in the receipts. This bounded legacy-comparable curve stops at 4K; the separately classified formal 8K screen is not silently mixed into it.", "evidence": "results/qwen38-flash-next-fp8-b70/README.md" } ], "estimates": [ { "id": "qwen38-flash-next-fp8-tp4-mtp0-context24k-estimate-v1", "state": "estimated", "selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 24576, "graph_mode": "off", "kv": "auto", "modality": "text" }, "metric": "decode_tok_s", "unit": "tok/s", "value": 3.32691, "interval": { "low": 1.663455, "high": 4.990364 }, "engine": { "name": "qwen38-flash-next-mtp0-context-estimator", "version": "1.0.0", "snapshot_sha256": "bc903a7ee9e638a06fb700bb107d8ff1ad11cfe104c8d351dfefbe27620ff02f" }, "generated_at": "2026-08-28T18:00:00Z", "basis_measurement_ids": [ "qwen38-flash-next-fp8-tp4-context4k-a1", "qwen38-flash-next-fp8-tp4-context8k-a1" ], "record": "data/qwen38-flash-next-fp8-tp4-mtp0-context-estimate-v1.json", "not_for_promotion": true, "limitations": "Grade-D legacy-runtime extrapolation only. No 24K boot, fit, completion, quality, speed, current-runtime, graph, MTP, TP1/TP2, vision, deployment, record, or promotion claim." }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-context32k-estimate-v1", "state": "estimated", "selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 32768, "graph_mode": "off", "kv": "auto", "modality": "text" }, "metric": "decode_tok_s", "unit": "tok/s", "value": 3.174425, "interval": { "low": 1.587212, "high": 4.761637 }, "engine": { "name": "qwen38-flash-next-mtp0-context-estimator", "version": "1.0.0", "snapshot_sha256": "bc903a7ee9e638a06fb700bb107d8ff1ad11cfe104c8d351dfefbe27620ff02f" }, "generated_at": "2026-08-28T18:00:00Z", "basis_measurement_ids": [ "qwen38-flash-next-fp8-tp4-context4k-a1", "qwen38-flash-next-fp8-tp4-context8k-a1" ], "record": "data/qwen38-flash-next-fp8-tp4-mtp0-context-estimate-v1.json", "not_for_promotion": true, "limitations": "Grade-D legacy-runtime extrapolation only. No 32K boot, fit, completion, quality, speed, current-runtime, graph, MTP, TP1/TP2, vision, deployment, record, or promotion claim." } ], "packets": [ { "id": "qwen38-flash-next-fp8-tp4-mtp1-placement-b70-32tps-20260906", "label": "Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1 + never-routed experts host-placed", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 005dc57895896f770157ea94f68e473e7447139e (placement overlay on 1b2a17c1) + staged kernels 2f829747 + public oneCCL 4ceafd1", "cards": 4, "status": "candidate", "evidence_level": "B70-verified originating-host replay", "coverage": [ "decode", "exactness", "quality", "runtime provenance", "recipe" ], "manifest": "packages/qwen38-flash-next-fp8-tp4-mtp1-placement-b70-32tps-20260906/package.json", "grades": { "capability": { "grade": "B", "scope": "user-facing TP4/EP4 deterministic full-decode-graph text serving with lossless MTP1 at configured max 4,352, one sequence, on the fixed cold realistic suite", "basis": "The certified line answers the 12-prompt realistic suite with every output identical to the no-speculation line, passed the quality profiles, and reproduced its exact-2K and exact-4K authorities on a separate server; a workload that routes to a host-placed expert pays a PCIe read for that row; long-context, concurrency, and clean-host installation are not covered.", "reviewed_at": "2026-09-06", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a226-promotion-attestation.json" ] }, "evidence": { "grade": "B", "scope": "exact TP4/EP4 graph lossless-MTP1 placement record at 31.929 tok/s class-balanced (A226) with the A225 battery on a separate server", "basis": "Model, placement overlay commit, hosted kernel stage, hosted oneCCL build, tuned map, placement file and frozen packet are pinned by hash; the record run retained every prompt and output SHA-256 (equal to the approved MTP0 and MTP1 rows), cache-zero fresh-response gates and the tuned-map selection receipt; a separate server reproduced both exact-depth authorities bit-for-bit. Only the originating host has replayed it.", "reviewed_at": "2026-09-06", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a226-realistic-suite-v1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a225-fresh-repeat-deterministic-summary.json" ] } } }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-hctriton-b70-37tps-20260907", "label": "Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1 + never-routed experts host-placed + Triton HC glue", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 622191221475b53cc6f7f4d847860939f4c300ab (Triton-HC overlay on the placement head 005dc578) + staged kernels 2f829747 + public oneCCL 4ceafd1", "cards": 4, "status": "candidate", "evidence_level": "B70-verified originating-host replay", "coverage": [ "decode", "exactness", "quality", "runtime provenance", "recipe" ], "manifest": "packages/qwen38-flash-next-fp8-tp4-mtp1-hctriton-b70-37tps-20260907/package.json", "grades": { "capability": { "grade": "B", "scope": "user-facing TP4/EP4 deterministic full-decode-graph text serving with lossless MTP1 at configured max 4,352, one sequence, on the fixed cold realistic suite", "basis": "The certified line answers the 12-prompt realistic suite deterministically with the quality profile of the certified rows (byte-identical exact-case outputs, one repeat hash, exact needle), reproduced its exact-2K and exact-4K pins on five servers, and keeps MTP1 lossless within the lineage; its outputs are a new authority (the Triton hyper-connection glue rounds at the last bf16 bit) rather than bit-identical to the torch-fallback rows; a workload that routes to a host-placed expert pays a PCIe read for that row; long-context, concurrency, and clean-host installation are not covered.", "reviewed_at": "2026-09-07", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp1-a272-promotion-attestation.json" ] }, "evidence": { "grade": "B", "scope": "exact TP4/EP4 graph lossless-MTP1 placement record at 31.929 tok/s class-balanced (A226) with the A225 battery on a separate server", "basis": "Model, placement overlay commit, hosted kernel stage, hosted oneCCL build, tuned map, placement file and frozen packet are pinned by hash; the record run retained every prompt and output SHA-256 (equal to the approved MTP0 and MTP1 rows), cache-zero fresh-response gates and the tuned-map selection receipt; a separate server reproduced both exact-depth authorities bit-for-bit. Only the originating host has replayed it.", "reviewed_at": "2026-09-06", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a226-realistic-suite-v1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a225-fresh-repeat-deterministic-summary.json" ] } } }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-qsafused-b70-38tps-20260907", "label": "Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1 + never-routed experts host-placed + both reference Triton kernels", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 6d8724577dabbee5fa0bbc70c4d927c6174c8d8a (fused-QSA overlay on the Triton-HC head 62219122) + staged kernels 2f829747 + public oneCCL 4ceafd1", "cards": 4, "status": "candidate", "evidence_level": "B70-verified originating-host replay", "coverage": [ "decode", "exactness", "quality", "runtime provenance", "recipe" ], "manifest": "packages/qwen38-flash-next-fp8-tp4-mtp1-qsafused-b70-38tps-20260907/package.json", "grades": { "capability": { "grade": "B", "scope": "user-facing TP4/EP4 deterministic full-decode-graph text serving with lossless MTP1 at configured max 4,352, one sequence, on the fixed cold realistic suite", "basis": "The certified line answers the 12-prompt realistic suite deterministically with the quality profile of the certified rows preserved byte for byte (seven of seven exact-case outputs identical, one repeat hash, exact needle), reproduced its pins across servers at both depths, and keeps MTP1 lossless within the lineage; its outputs are a new authority whose difference is at depth (exact-2K coincides with the certified stream, exact-4K does not), so a consumer pinning 4K outputs must re-pin; a workload that routes to a host-placed expert pays a PCIe read for that row; long-context, concurrency, and clean-host installation are not covered.", "reviewed_at": "2026-09-07", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260907-tp4-mtp1-a306-promotion-attestation.json" ] }, "evidence": { "grade": "B", "scope": "exact TP4/EP4 graph lossless-MTP1 placement record at 31.929 tok/s class-balanced (A226) with the A225 battery on a separate server", "basis": "Model, placement overlay commit, hosted kernel stage, hosted oneCCL build, tuned map, placement file and frozen packet are pinned by hash; the record run retained every prompt and output SHA-256 (equal to the approved MTP0 and MTP1 rows), cache-zero fresh-response gates and the tuned-map selection receipt; a separate server reproduced both exact-depth authorities bit-for-bit. Only the originating host has replayed it.", "reviewed_at": "2026-09-06", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a226-realistic-suite-v1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260906-tp4-mtp1-a225-fresh-repeat-deterministic-summary.json" ] } } }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-w13n64-b70-34tps-20260908", "label": "Qwen3.8 Flash-Next FP8 ? TP4+EP4 graph, no speculation, W13-N64", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 2a372e860e273273357cb7437ac7de1694304f9f + W13-N64 tuned map", "cards": 4, "status": "candidate", "evidence_level": "B70-verified originating-host replay; clean-host replay pending", "coverage": [ "decode", "exactness", "runtime provenance", "recipe" ], "manifest": "packages/qwen38-flash-next-fp8-tp4-mtp0-w13n64-b70-34tps-20260908/package.json", "grades": { "capability": { "grade": "B", "scope": "Originating-host TP4/EP4 no-speculation W13-N64 text serving on the fixed cold realistic suite", "basis": "The existing promotion attestation binds the A326 fresh-server repeat to unchanged MTP0 output identity at exact 2K and 4K across three servers. Clean-host installation, container delivery, and a context sweep remain pending.", "reviewed_at": "2026-09-08", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260908-tp4-mtp0-a326-w13n64-promotion-attestation.json" ] }, "evidence": { "grade": "B", "scope": "A325 and A326 W13-N64 MTP0 cold-suite runs, with A326 retained as the lower repeat", "basis": "The package identity pins the performance evidence and three-server exact-depth output comparisons by hash. The recorded medians are 34.510128 and 34.495292 tok/s; this is originating-host evidence only.", "reviewed_at": "2026-09-08", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260908-tp4-mtp0-a326-w13n64-promotion-attestation.json", "repro/qwen38-flash-next-fp8-tp4-mtp0-w13n64-b70-34tps-20260908/identity.json" ] } } }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-lossless-b70-27tps-20260905", "label": "Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1b2a17c1e7c41985d6a5e0eb324ada4775c25e60 + staged kernels 2f829747 + public oneCCL 4ceafd1", "cards": 4, "status": "candidate", "evidence_level": "B70-verified originating-host replay", "coverage": [ "decode", "exactness", "quality", "runtime provenance", "recipe" ], "manifest": "packages/qwen38-flash-next-fp8-tp4-mtp1-lossless-b70-27tps-20260905/package.json", "grades": { "capability": { "grade": "B", "scope": "user-facing TP4/EP4 deterministic full-decode-graph text serving with lossless MTP1 at configured max 4,352, one sequence, on the fixed cold realistic suite", "basis": "The certified line answers the 12-prompt realistic suite with every output identical to the no-speculation line, passed the direct-answer (6/7) and official-thinking (25/25) quality profiles, and reproduced its exact-2K and exact-4K authorities on a fresh server; long-context, concurrency, and clean-host installation are not covered, so it is a candidate package rather than a starter guide.", "reviewed_at": "2026-09-06", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260905-tp4-mtp1-a189-promotion-attestation.json" ] }, "evidence": { "grade": "B", "scope": "exact TP4/EP4 graph MTP1 record at 27.048 tok/s class-balanced (A189) with the A190 fresh-server frozen-client repeat", "basis": "Model, overlay commit, hosted kernel stage, hosted oneCCL build, tuned map, and frozen packet are pinned by hash; the record run retained every prompt and output SHA-256, cache-zero fresh-response gates, and the exactness verifier; a second server reproduced the exact-depth authorities bit-for-bit. Only the originating host has replayed it.", "reviewed_at": "2026-09-06", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/data/20260905-tp4-mtp1-a189-realistic-suite-v1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260905-tp4-mtp1-a190-fresh-repeat-deterministic-summary.json" ] } } }, { "id": "qwen38-flash-next-fp8-tp4-research", "label": "Qwen3.8 Flash-Next FP8 · TP4 research server", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "cards": 4, "status": "research-only-quality-caveat", "evidence_level": "B70-screened", "coverage": [ "TP4/EP4", "MTP0", "eager", "text-only", "configured max 512/1536/3072/4352/8448", "short quality plus screened 1K/2K/4K and formal 8K context" ], "grades": { "capability": { "grade": "C", "scope": "user-facing TP4 eager MTP0 text-serving readiness on the retained legacy runtime through exact active 8K", "basis": "The packet boots, serves, repeats, and passes exact-depth gates through 8K, but direct-answer quality is 6/7 semantically and the 8K comparison sequence is incomplete. A separate packet now carries the additive current-runtime short/4K replay; vision, clean-host stability, and deployment sealing remain open.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-8448-context-screen.json" ] }, "evidence": { "grade": "C", "scope": "exact TP4/EP4 eager MTP0 text-serving source/runtime through the formal 8K context screen", "basis": "Exact model, source overlay, staged runtime, serving, short quality, a 1K needle, the 12-prompt realistic suite, and exact-1K observations are retained. The original 2K stop-gate evidence remains visible; protocol-v2 2K and additive 4K arms passed fixed-set repeats, exact cache-zero gates, and three exact-depth legacy-comparable observations each. The 8K arm passed incremental quality and formal cache-zero gates but failed the required third legacy comparison, so it adds formal capability evidence with an explicit stability caveat rather than a legacy curve point. Full quality, 16K+, vision and clean-host replay remain missing.", "reviewed_at": "2026-08-27", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-attempt19-production-qualification.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-1536-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-3072-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-3072-context-repeat-v2-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-4352-context-screen.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp0-8448-context-screen.json" ] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "qwen38-flash-next-fp8-tp4-attempt19", "sample_index": 0, "value": 5.221849709057954, "unit": "tok/s", "workload": "One instrumentation-free server; p146/o256/c1; three sequential repetitive-prompt requests; full-output rate after first text, not the conventional 99-interval final gate", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-attempt19-production-qualification.json" }, "manifest": "results/qwen38-flash-next-fp8-b70/README.md" }, { "id": "qwen38-flash-next-fp8-tp4-mtp0-current-research", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP0 current-runtime research screen", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "cards": 4, "status": "research-only-quality-caveat", "evidence_level": "B70-screened", "coverage": [ "TP4/EP4", "MTP0", "eager", "text-only", "configured max 4352", "short screen plus repeated exact active-4K screen", "same-boot quality and card-clean teardown" ], "grades": { "capability": { "grade": "C", "scope": "user-facing TP4 eager MTP0 current-runtime text-serving readiness at short context and exact active 4K", "basis": "The exact current source and staged runtime booted, served, passed the recovery canary, completed the short and exact-4K screens, and returned all cards idle. Direct-answer quality remains 6/7 and fresh-server, clean-host, graph, vision, deeper-context, and deployment sealing remain open.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-runtime-anchor-attempt4-result.json" ] }, "evidence": { "grade": "C", "scope": "exact current-runtime TP4/EP4 eager MTP0 short and active-4096 text screens under configured max 4352", "basis": "The receipt pins model, vLLM, XPU-kernel source, staged runtime, topology, placement, cache, commands, artifact hashes, and postflight. The run passed a fresh four-rank preflight, cache-zero recovery canary, 6/7 semantic battery, 16/16 repeat, exact cache-zero 4K needle, three established short rows, and two exact p4096/o128 rows with one output-token hash. The short harness lacks per-row cache and finish fields, and the 4K repeats are same-boot only; neither clean-host nor deployment qualification is claimed.", "reviewed_at": "2026-08-28", "evidence": [ "experiments/qwen38-flash-next-fp8-b70/notes/2026-08-28-tp4-mtp0-current-runtime-anchor-a4-result.md", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-runtime-anchor-attempt4-result.json" ] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp0-current-context4k-a4", "sample_index": 0, "value": 4.7578181021380175, "unit": "tok/s", "workload": "Two exact p4096/o128 requests; conventional decode across the 99 inter-token intervals between generated-token events 1 and 100 after TTFT", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-runtime-anchor-attempt4-result.json" }, "manifest": "results/qwen38-flash-next-fp8-b70/README.md" }, { "id": "qwen38-flash-next-fp8-tp4-mtp1-research", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP1 research screen", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "cards": 4, "status": "research-only-quality-caveat", "evidence_level": "B70-screened", "coverage": [ "TP4/EP4", "MTP1", "eager", "text-only", "configured max 512 and 4352/headroom32", "configured max 1536 and 3072 retained as first-request bounded negatives", "configured max 8448 retained as an active-8K cross-runtime parity quarantine", "matched MTP0 quality plus 16/16 repeat and exact-4K needle", "three exact-4K deployment-shaped rows plus a separate formal row" ], "grades": { "capability": { "grade": "D", "scope": "user-facing TP4 eager MTP1 text-serving readiness across the packet's declared short and context cells", "basis": "Configured-512 and exact-4K serving are useful screened evidence, but active 1K/2K did not complete and active 8K diverged from the cross-runtime authority. The packet is not deployment-ready even though its passing subcells remain valid.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-8448-context-attempt1-parity-quarantine.json" ] }, "evidence": { "grade": "C", "scope": "exact TP4/EP4 eager MTP1 text serving at configured max 512 and configured max 4352 with a 32-block cache allocation, including active 4096-token context; configured max 1536, 3072, and 8448 are retained separately as active-1K/2K bounded negatives and an active-8K cross-runtime parity quarantine", "basis": "The exact model, MTP adapter, preserved staged runtime, healthy API, all 26 MTP0 baseline comparisons, 16/16 repeat stability, cache-zero needles, complete usage, positive draft acceptance, short speed rows, and the exact-4K formal plus three-row deployment-shaped screen are retained. At 4K, the median was 8.904 tok/s decode, 232.1 seconds TTFT, and 0.981 tok/s wall output with 528/539 cumulative draft acceptance. Separate configured-1536 and configured-3072 boots stopped during their first active-1K and active-2K requests without returning output tokens. The active-8K arm completed all 25 generic gates with zero cache reuse and positive MTP1 acceptance, but diverged from the frozen cross-runtime MTP0 authority at generated token 72; its 4.151 tok/s and 953.3-second TTFT remain diagnostic only. None of these quarantines changes the passing short or exact-4K evidence or grants performance, quality, or deployment credit. The inherited target quality remains 5/7; the 32-block recipe is not a minimum-cache or causal cache claim, and vision, clean-boot replay, and deployment qualification remain missing. MTP3 remains the preferred exact-4K recipe.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-512-attempt3-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-1536-context-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-3072-context-attempt2-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-8448-context-attempt1-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-4352-headroom32-attempt1-result.json" ] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp1-a3", "sample_index": 0, "value": 9.37225436776222, "unit": "tok/s", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp1-512-attempt3-result.json" }, "manifest": "results/qwen38-flash-next-fp8-b70/README.md" }, { "id": "qwen38-flash-next-fp8-tp4-mtp3-research", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP3 research screen", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "cards": 4, "status": "research-only-quality-caveat", "evidence_level": "B70-screened", "coverage": [ "TP4/EP4", "MTP3", "eager", "text-only", "configured max 512 and 4352", "configured max 1536 retained as an active-1K external-signal quarantine", "configured max 3072 retained as an active-2K target-parity quarantine", "configured max 8448 retained as an active-8K bounded no-receipt quarantine", "matched MTP0 quality plus 16/16 repeat and exact-4K needle", "variable 512-token screen plus exact-4K service and formal rows" ], "grades": { "capability": { "grade": "C", "scope": "user-facing TP4 eager MTP3 text-serving readiness, centered on the preferred exact-4K recipe", "basis": "Exact-4K capability, parity, acceptance, and the strongest qualified 4K service rate make this the preferred current-runtime recipe. Active 1K/2K/8K remain quarantined and repeated-session, vision, clean-host, and deployment sealing are incomplete.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-4352-attempt1-result.json" ] }, "evidence": { "grade": "C", "scope": "exact TP4/EP4 eager MTP3 text serving at configured max 512 and 4352, including active 4096-token context; configured max 1536, 3072, and 8448 are retained separately as active-1024 external-signal, active-2048 target-parity, and active-8192 bounded no-receipt quarantines", "basis": "The exact model, MTP adapter, preserved staged runtime, fixed cache allocations, healthy API, all 26 MTP0 baseline comparisons, 16/16 repeat stability, exact cache-zero 4K needle and formal gate, complete usage, cumulative draft engagement, and both short and exact-4K speed rows are retained. The inherited target quality remains 5/7; the 4K service median pairs 15.502 tok/s decode with 187.9-second TTFT and 1.246 tok/s wall output. A separate active-1K boot received an external SIGTERM before completion, and a separate active-2K request completed but diverged from the cross-runtime MTP0 authority at generated token five. The exact 32-block active-8K boot reported 9,654 cache tokens but its sole request reached the fixed 900-second bound without a completed receipt or durably recorded output token. None of these quarantines changes the passing short or 4K evidence or grants speed, quality, or deployment credit. The official-thinking transfer completed 19 of 25 required rows; all completed rows passed, but repeated-session stability, fresh-boot replay, deployment, and vision remain missing.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-512-attempt4-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-1536-context-attempt1-external-stop.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-3072-context-attempt1-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-8448-context-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-4352-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-official-quality-attempt2-result.json" ] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp3-context4k-a1", "sample_index": 0, "value": 15.50156510641242, "unit": "tok/s", "workload": "Three separately salted exact p4096/o256/c1 requests with no harness-added warmups; full-output rate after first text; formal p4096/o128 99-interval row retained separately", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp3-4352-attempt1-result.json" }, "manifest": "results/qwen38-flash-next-fp8-b70/README.md" }, { "id": "qwen38-flash-next-fp8-tp4-mtp2-research", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP2 research screen", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "cards": 4, "status": "research-only-quality-caveat", "evidence_level": "B70-screened", "coverage": [ "TP4/EP4", "MTP2", "eager", "text-only", "configured max 512 and 4352/headroom32", "configured max 1536 retained as an active-1K clean-host quarantine", "configured max 3072 and 8448 retained as active-2K and active-8K target-parity quarantines", "matched MTP0 quality plus 16/16 repeat and exact-4K needle", "short and exact-4K three-row speed screens plus separate formal rows" ], "grades": { "capability": { "grade": "D", "scope": "user-facing TP4 eager MTP2 text-serving readiness across the packet's declared short and context cells", "basis": "Configured-512 and exact-4K subcells serve, but 1K is clean-host quarantined, 2K/8K diverge from the authority, and both bounded 16K treatments produced no output with the final postflight failing. No deep-context deployment path is currently open.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-16512-scheduler32-attempt2-runtime-timeout-quarantine.json" ] }, "evidence": { "grade": "C", "scope": "exact TP4/EP4 eager MTP2 text serving at configured max 512 and configured 4352/headroom32; configured max 1536, 3072, and 8448 are retained separately as active-1024 clean-host and active-2048/8192 target-parity quarantines", "basis": "The exact model, MTP adapter, preserved staged runtime, healthy APIs, all 26 MTP0 baseline comparisons, 16/16 repeat stability, cache-zero needles, complete usage, positive draft acceptance, and three target-hash rows at both configured 512 and exact 4K are retained. At exact 4K, the 32-block recipe completed the formal and three-row gates at 9.893 tok/s decode, 263.3 seconds TTFT, and 0.891 tok/s wall output. A separate active-1K boot returned the exact frozen MTP0 text twice, with zero cache reuse and perfect acceptance at both MTP2 positions, but 11 corrected local-NVMe events after the frozen journal cutoff failed its clean-host gate; its 10.683 and 12.642 tok/s observations are diagnostic only. Separate active-2K and active-8K requests completed their generic depth gates with MTP2 active but diverged from the frozen MTP0 token arrays at generated tokens 13 and 27, so their observed 4.527 and 6.235 tok/s also receive no speed or quality credit. The 8K arm's 22 corrected storage/root-port records separately block clean-host wording; neither arm recorded a B70 event. The prior 21-block stop remains disclosed; none of these arms proves a speed gain, minimum cache, or universal causality. Inherited quality remains 5/7, while vision, clean-host replay, and deployment qualification remain missing. MTP3 remains preferred at exact 4K.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-512-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-1536-context-attempt1-host-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-3072-context-attempt1-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-8448-context-attempt2-parity-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-4352-headroom32-attempt2-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-4352-attempt1-mixed-quarantine.json" ] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp2-a1", "sample_index": 0, "value": 11.895061402541456, "unit": "tok/s", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup per invocation then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp2-512-attempt1-result.json" }, "manifest": "results/qwen38-flash-next-fp8-b70/README.md" }, { "id": "qwen38-flash-next-fp8-tp4-mtp4-research", "label": "Qwen3.8 Flash-Next FP8 · TP4 MTP4 research screen", "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "quantization": "FP8 block-128", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "cards": 4, "status": "research-only-quality-caveat", "evidence_level": "B70-screened", "coverage": [ "TP4/EP4", "MTP4", "eager", "text-only", "configured max 512", "configured max 1536 retained as an active-1K teardown quarantine", "configured max 3072 retained as an active-2K no-output and reset quarantine", "configured max 4352 retained as an exact-4K runtime quarantine", "configured max 8448 retained as an exact active-8K cross-runtime parity quarantine", "matched MTP0 quality plus 16/16 repeat and small needle", "three-row speed screen" ], "grades": { "capability": { "grade": "D", "scope": "user-facing TP4 eager MTP4 text-serving readiness across the packet's declared short and context cells", "basis": "The configured-512 screen is fast and fully engages all four draft positions, but every tested deeper context is quarantined for lifecycle, completion, parity, or clean-host reasons. The 20.727 tok/s screen remains valid evidence, not deployment readiness.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt2-parity-quarantine.json" ] }, "evidence": { "grade": "C", "scope": "exact TP4/EP4 eager MTP4 text serving at configured max 512; configured max 1536, 3072, 4352, and 8448 are retained separately as active-1K teardown, active-2K no-output/reset, exact-4K runtime, and exact active-8K cross-runtime parity quarantines", "basis": "The exact model, MTP adapter, preserved staged runtime, exact 24-block fixed allocation, healthy API, all 26 MTP0 baseline comparisons, 16/16 repeat stability, small cache-zero needle, complete usage, 1,716/1,716 cumulative draft acceptance, and three corrected target-hash speed rows are retained. A separate 29-block active-1K boot returned the frozen MTP0 hash twice with zero cache reuse and perfect 204/204 acceptance across all four positions on each request, but its detached supervisor did not forward the exact stop to the server group. Direct recovery shut it down cleanly, yet the frozen teardown rule gives the cell no speed, quality, or deployment credit. The successor active-2K boot passed every startup gate but returned no receipt or output before the bounded client and engine timeouts; its teardown window recorded compute- and copy-class resets on all four cards. The exact active-8K boot then passed all generic request and four-position MTP4 mechanism gates, but first diverged from the frozen cross-runtime/cache MTP0 authority at generated token 27. Its 4.026 tok/s observation is diagnostic only and does not isolate MTP4 as the cause. The first active-8K launch was a superseded pre-request command-identity stop with no matrix credit. Corrected-only local-NVMe records block clean-host qualification in these context arms. The inherited target quality remains 5/7; exact-4K MTP4 is separately quarantined, while vision, clean-host replay, and deployment qualification remain missing.", "reviewed_at": "2026-08-28", "evidence": [ "results/qwen38-flash-next-fp8-b70/README.md", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-512-attempt1-result.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-1536-context-attempt1-teardown-quarantine.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-3072-context-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-4352-attempt1-bounded-negative.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt1-pre-request-stop.json", "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt2-parity-quarantine.json" ] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "qwen38-flash-next-fp8-tp4-mtp4-a1", "sample_index": 0, "value": 20.72717637199404, "unit": "tok/s", "workload": "One instrumentation-free server; p146/o256/c1; one untimed warmup per invocation then three sequential measured requests; full-output rate after first text, not the conventional 99-interval final gate", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-512-attempt1-result.json" }, "manifest": "results/qwen38-flash-next-fp8-b70/README.md" } ], "views": [ { "id": "qwen-flash-next-tp-screen", "title": "Research speed screen by TP", "subtitle": "Only TP4 has bounded screens; each MTP cell has three same-server samples under the declared workload, so no TP mini-curve is implied", "x_label": "tensor parallel cards", "discrete": true, "missing_x": [ 1, 2 ], "metrics": [ "decode_tok_s", "ttft_ms" ], "series": [ { "label": "official FP8 · eager MTP0", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-attempt19" ], "x_from": "config.tp" }, { "label": "official FP8 · eager MTP1", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp1-a3" ], "x_from": "config.tp" }, { "label": "official FP8 · eager MTP2 (variable screen)", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp2-a1" ], "x_from": "config.tp" }, { "label": "official FP8 · eager MTP3 (variable screen)", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp3-a4" ], "x_from": "config.tp" }, { "label": "official FP8 · eager MTP4", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp4-a1" ], "x_from": "config.tp" } ] }, { "id": "qwen-flash-next-context-screen", "title": "Legacy-comparable after-first-text context screen", "subtitle": "The plotted 5.22/5.13/5.23/5.23 tok/s values use the legacy after-first-text comparison and stop at 4K. Separate formal cache-zero rates are 3.865 at 2K, 4.456 at 4K, and 3.980 tok/s at 8K; 8K is not plotted here because its third comparison row did not complete. All cells retain the short-quality caveat.", "x_label": "active context tokens", "metrics": [ "decode_tok_s", "ttft_ms" ], "series": [ { "label": "official FP8 · TP4 eager MTP0", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-context-screen" ], "x_from": "active_context_tokens" } ] }, { "id": "qwen-flash-next-context4k-mtp-tradeoff", "title": "Exact-4K service tradeoff by MTP depth", "subtitle": "Workload-aligned p4096/o256/c1 service screens with no harness-added warmups, but from different vLLM source revisions and cache allocations. MTP1 and MTP2 use 32-block headroom recipes; MTP3 remains faster with lower TTFT and higher wall output. Separate single-point series avoid implying a causal MTP-depth or cache curve; only MTP4 remains quarantined.", "x_label": "MTP depth", "discrete": true, "missing_x": [ 4 ], "metrics": [ "decode_tok_s", "wall_output_tok_s", "ttft_ms" ], "series": [ { "label": "official FP8 · TP4 eager MTP0", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-context4k-service-a1" ], "x_from": "config.mtp" }, { "label": "official FP8 · TP4 eager MTP1 · headroom32", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp1-context4k-headroom32-a1" ], "x_from": "config.mtp" }, { "label": "official FP8 · TP4 eager MTP2 · headroom32", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp2-context4k-headroom32-a2" ], "x_from": "config.mtp" }, { "label": "official FP8 · TP4 eager MTP3", "measurement_ids": [ "qwen38-flash-next-fp8-tp4-mtp3-context4k-a1" ], "x_from": "config.mtp" } ] } ], "coverage_views": [ { "id": "qwen-flash-next-tp4-mtp-by-context", "label": "Practical TP4 eager text coverage", "fixed": "Official FP8 child artifact on TP4/EP4, eager text, and automatic KV. MTP0 uses vLLM 658965050 plus kernels 2f829747; MTP1-4 use vLLM 1372c62d plus staged kernels 2f829747. This is a coverage map across exact runtime identities, not a controlled depth curve.", "row_axis": { "key": "mtp", "label": "MTP depth", "prefix": "MTP" }, "column_axis": { "key": "active_context_tokens", "label": "Active context", "prefix": "", "value_labels": { "0": "0", "1024": "1K", "2048": "2K", "4096": "4K", "8192": "8K" } }, "fixed_selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "runtime_family": "vLLM XPU", "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "graph_mode": "off", "kv": "auto", "modality": "text" }, "rows": [ 0, 1, 2, 3, 4 ], "columns": [ 0, 1024, 2048, 4096, 8192 ], "cells": { "0:0": { "state": "lab-screened", "label": "5.222 tok/s; quality caveat", "evidence_id": "qwen38-flash-next-fp8-tp4-attempt19", "packet_id": "qwen38-flash-next-fp8-tp4-research", "selectors": { "runtime": "vLLM XPU 658965050 + kernels 2f829747" } }, "0:1024": { "state": "lab-screened", "label": "5.134 tok/s exact-1K screen", "evidence_id": "qwen38-flash-next-fp8-tp4-context1k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-research", "selectors": { "runtime": "vLLM XPU 658965050 + kernels 2f829747" } }, "0:2048": { "state": "lab-screened", "label": "5.228 tok/s exact-2K protocol-v2 screen", "evidence_id": "qwen38-flash-next-fp8-tp4-context2k-v2-a2", "packet_id": "qwen38-flash-next-fp8-tp4-research", "selectors": { "runtime": "vLLM XPU 658965050 + kernels 2f829747" } }, "0:4096": { "state": "lab-screened", "label": "4.456 tok/s formal exact-4K", "evidence_id": "qwen38-flash-next-fp8-tp4-context4k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-research", "selectors": { "runtime": "vLLM XPU 658965050 + kernels 2f829747" } }, "0:8192": { "state": "lab-screened", "label": "3.980 tok/s formal exact-8K", "evidence_id": "qwen38-flash-next-fp8-tp4-context8k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-research", "selectors": { "runtime": "vLLM XPU 658965050 + kernels 2f829747" } }, "1:0": { "state": "lab-screened", "label": "9.372 tok/s; matched quality caveat", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp1-a3", "packet_id": "qwen38-flash-next-fp8-tp4-mtp1-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "1:1024": { "state": "quarantined", "label": "768 computed prompt tokens; no output", "reason": "The exact 32-block configured-1536 identity passed source, runtime, four-rank, placement, cache, capacity, and health gates, then reached the unchanged worker-response deadline during its first 1K request. No output was returned; request two was not sent and the separate 2K boot did not run.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-1536-context-attempt1-bounded-negative.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "1:2048": { "state": "quarantined", "label": "360-second client timeout; 448 computed tokens; no output or speed", "reason": "The exact 32-block configured-3072 identity passed all startup gates and exposed 7,561 cache tokens. Its first exact-2K exchange had a zero-byte completion body and no output token recorded when the fixed 360-second client bound expired; the subsequent engine diagnostic showed 448 computed prompt tokens and zero output, while vLLM completed-request, token, and MTP counters remained zero. The engine independently reported its own sampling timeout. Request two was not sent. The post-failure teardown window recorded compute- and copy-class resets on all four cards before all four were rediscovered; no post-reset collective was run. Existing MTP1 configured-512 and exact-4K passes and every captured speed remain unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-3072-context-attempt2-bounded-negative.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "1:4096": { "state": "lab-screened", "label": "8.904 tok/s decode; headroom32", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp1-context4k-headroom32-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp1-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "1:8192": { "state": "quarantined", "label": "exact 8K completed; parity mismatch at token 73; 4.151 tok/s diagnostic only", "reason": "The exact 32-block current-source identity exposed 13,516 cache tokens and completed one p8192/o128 request with exact usage, zero cache reuse, all 25 generic depth gates, and positive MTP1 counters at position zero. Its token array first diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 72. The 4.151 tok/s rate and 953.3-second TTFT receive no speed or quality credit. Controlled teardown passed its cleanup gates and returned all four cards idle; the intentional stop retained the known shutdown-time output-handler notice and one shared-memory cleanup warning. The host window contained corrected local-NVMe events but no B70 event.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-8448-context-attempt1-parity-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "2:0": { "state": "lab-screened", "label": "11.895 tok/s; variable matched-quality screen", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp2-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp2-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "2:1024": { "state": "quarantined", "label": "exact parity twice; corrected local-NVMe events; diagnostic only", "reason": "Both exact active-1K requests returned the frozen MTP0 text hash with zero cache reuse, identical text, and perfect MTP2 acceptance at positions zero and one. Request one observed 10.683 tok/s after first text and the repeat sentinel 12.642 tok/s. Eleven corrected local-NVMe events after the preregistered journal cutoff failed the strict clean-host gate, so neither rate receives speed, quality, or deployment credit. No event named a B70 address; existing MTP2 configured-512, active-2K, and exact-4K results remain unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-1536-context-attempt1-host-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "2:2048": { "state": "quarantined", "label": "completed; target-parity mismatch at token 13; no speed credit", "reason": "The exact 32-block configured-3072 identity passed all startup gates. Its first exact-2K request returned 128 tokens with zero cache reuse and active MTP2 counters, but the token array diverged from the frozen MTP0 authority at zero-based generated-token index 12. Request two was not sent. The observed 4.527 tok/s is diagnostic only and does not alter the passing MTP2 512 or exact-4K results.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-3072-context-attempt1-parity-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "2:4096": { "state": "lab-screened", "label": "9.893 tok/s decode; headroom32", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp2-context4k-headroom32-a2", "packet_id": "qwen38-flash-next-fp8-tp4-mtp2-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "2:8192": { "state": "quarantined", "label": "exact 8K completed; parity mismatch at token 27; 6.235 tok/s diagnostic only", "reason": "The exact 32-block current-source identity exposed 11,264 cache tokens and completed one p8192/o128 request with exact usage, zero cache reuse, all generic depth gates, and positive MTP2 counters at both positions. Its token array first diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 26. The 6.235 tok/s rate and 649.7-second TTFT receive no speed or quality credit. The bounded host window contained corrected storage/root-port events but no B70 event; teardown was clean and all four cards returned idle.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-8448-context-attempt2-parity-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "3:0": { "state": "lab-screened", "label": "14.889 tok/s; variable matched-quality screen", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp3-a4", "packet_id": "qwen38-flash-next-fp8-tp4-mtp3-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "3:1024": { "state": "quarantined", "label": "external SIGTERM during request one; six accepted draft tokens; no completed response or speed", "reason": "The exact 25-block local-NVMe identity passed source, runtime, fresh four-rank, placement, cache, capacity, served-identity, and health gates. Request one began under the frozen protocol, then the server received an external SIGTERM before the response completed. No request JSON, usage, output hash, or performance result exists. Partial server metrics showed six drafted and six accepted tokens with 1.000 acceptance at all three MTP3 positions, but those counters receive no parity, speed, quality, or deployment credit. Request two was not sent. Existing MTP3 configured-512, active-2K, and exact-4K results remain unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-1536-context-attempt1-external-stop.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "3:2048": { "state": "quarantined", "label": "completed; target-parity mismatch at token five; no speed credit", "reason": "The exact 25-block configured-3072 identity passed source, runtime, four-rank, placement, cache, capacity, and health gates. Its first exact-2K request returned 128 tokens with zero cache reuse and active MTP3 counters, but the token array diverged from the frozen MTP0 authority at zero-based generated-token index 4. Request two was not sent. The observed 5.932 tok/s is diagnostic only and does not alter the passing MTP3 512 or exact-4K results.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-3072-context-attempt1-parity-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "3:4096": { "state": "lab-screened", "label": "15.502 tok/s decode; exact-4K", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp3-context4k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp3-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "3:8192": { "state": "quarantined", "label": "900-second client timeout; no completed receipt or speed", "reason": "The exact current-source TP4/EP4 eager MTP3 boot passed source/runtime, fresh four-rank, placement, served-identity, health, and 32-block capacity gates, reporting 9,654 cache tokens. Its sole p8192/o128 request reached the fixed 900-second client bound without a completed response receipt or any durably recorded output token. No usage, cache-zero, MTP-counter, parity, quality, TTFT, or speed result is claimed. The failed-request supervisor path cleanly removed the listener, process group, compile/RPC paths and returned all four cards idle. The host window contained corrected local-NVMe events but no B70-addressed event. Existing MTP3 configured-512, active-2K, exact-4K, and every captured speed remain unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-8448-context-attempt1-bounded-negative.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "4:0": { "state": "lab-screened", "label": "20.727 tok/s; matched quality caveat", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp4-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp4-research", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "4:1024": { "state": "quarantined", "label": "exact parity twice; detached teardown gate failed; diagnostic only", "reason": "Both exact active-1K requests returned the frozen MTP0 text hash with zero cache reuse, identical text, and perfect MTP4 acceptance at positions zero through three. Request one observed 13.326 tok/s after first text and the repeat sentinel 17.291 tok/s. The exact stop sentinel terminated the timeout supervisor without reaching the detached server group; direct recovery produced an orderly shutdown, but the frozen teardown rule quarantines the cell. Seven corrected-only local-NVMe records separately block clean-host qualification. Existing MTP4 configured-512 and exact-4K results remain unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-1536-context-attempt1-teardown-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "4:2048": { "state": "quarantined", "label": "360-second client timeout; 384 computed tokens; no output; four-card teardown resets", "reason": "The exact 29-block configured-3072 identity passed source, runtime, fresh four-rank, placement, cache, capacity, served-identity, and health gates. Request one reached the fixed 360-second client bound without a response receipt; about five seconds later the engine independently reported its own sampling timeout at 384 computed prompt tokens and zero output. No HTTP status, usage, parity hash, MTP counter delta, or speed exists, and request two was not sent. The corrected supervisor returned zero and left no listener, recorded process, compile path, or RPC path, but the teardown window recorded compute- and copy-class resets on all four B70s. All four cards were rediscovered at low memory use; no post-reset collective or known-good generation canary was run. Existing MTP4 configured-512, active-1K, exact-4K, and all captured speeds remain unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-3072-context-attempt1-bounded-negative.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "4:4096": { "state": "quarantined", "label": "worker timeout at 3,904 computed tokens; no speed", "reason": "The exact 29-block configuration admitted 4,352 tokens but stopped during the 4K quality request; no durable quality or timing result was produced", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-4352-attempt1-bounded-negative.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } }, "4:8192": { "state": "quarantined", "label": "exact 8K completed; parity mismatch at token 27; 4.026 tok/s diagnostic only", "reason": "The exact 36-block current-source identity exposed 9,504 cache tokens and completed one p8192/o128 request with exact usage, zero cache reuse, all 25 generic depth gates, and positive MTP4 counters at all four positions. Its token array first diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 26. The 4.026 tok/s rate and 918.4-second TTFT receive no speed or quality credit. Controlled teardown passed its cleanup gates and returned all four cards idle; the five-second grace expired before the remaining EngineCore and workers were stopped, and the intentional stop retained the known output-handler and shared-memory cleanup notices. Six corrected local-NVMe endpoint records block clean-host wording, but no B70 event appears. The superseded attempt-1 command-identity stop sent no request and grants no matrix credit.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt2-parity-quarantine.json", "selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747" } } } }, { "id": "qwen-flash-next-legacy-mtp0-context-estimates", "label": "Archived MTP0 context anchors and estimates", "fixed": "Legacy vLLM 658965050 TP4/EP4 eager-text MTP0 only. The 4K/8K cells are formal exact-depth measurements; 24K/32K are deterministic Grade-D extrapolations with 50%-150% bands. They do not transfer to the latest runtime or qualify boot, fit, quality, deployment, records, or promotion.", "row_axis": { "key": "mtp", "label": "MTP depth", "prefix": "MTP" }, "column_axis": { "key": "active_context_tokens", "label": "Active context", "prefix": "", "value_labels": { "4096": "4K", "8192": "8K", "24576": "24K", "32768": "32K" } }, "fixed_selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "graph_mode": "off", "kv": "auto", "modality": "text" }, "rows": [ 0 ], "columns": [ 4096, 8192, 24576, 32768 ], "cells": { "0:4096": { "state": "lab-screened", "label": "4.456 tok/s formal exact-4K", "evidence_id": "qwen38-flash-next-fp8-tp4-context4k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-research" }, "0:8192": { "state": "lab-screened", "label": "3.980 tok/s formal exact-8K", "evidence_id": "qwen38-flash-next-fp8-tp4-context8k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-research" }, "0:24576": { "state": "estimated", "label": "Grade-D legacy-runtime extrapolation only", "estimate_id": "qwen38-flash-next-fp8-tp4-mtp0-context24k-estimate-v1" }, "0:32768": { "state": "estimated", "label": "Grade-D legacy-runtime extrapolation only", "estimate_id": "qwen38-flash-next-fp8-tp4-mtp0-context32k-estimate-v1" } } }, { "id": "qwen-flash-next-tp4-deep-context", "label": "TP4 eager text deeper-context coverage", "fixed": "Official FP8 child artifact on current-source vLLM 1372c62d plus staged kernels 2f829747, TP4/EP4 eager text, and automatic KV. Single generic requests and bounded negatives are Grade-D matrix evidence until semantic and repeat qualification exists.", "row_axis": { "key": "mtp", "label": "MTP depth", "prefix": "MTP" }, "column_axis": { "key": "active_context_tokens", "label": "Active context", "prefix": "", "value_labels": { "16384": "16K", "24576": "24K", "32768": "32K" } }, "fixed_selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "graph_mode": "off", "kv": "auto", "modality": "text" }, "rows": [ 0, 1, 2, 3, 4 ], "columns": [ 16384, 24576, 32768 ], "cells": { "0:16384": { "state": "quarantined", "label": "16K sometimes completes; same-server corruption and fresh-server timeout; diagnostic only", "reason": "The current-source 33-block identity first completed one generic p16384/o128 request at a diagnostic-only 5.219 tok/s. The semantic program then returned one correct 16,213-token fresh response, corrupted the identical same-server repeat into repeated text despite zero reported cache use, and stopped a separate fresh-server attempt at 1,600 computed prompt tokens with no first output after an RPC sampling timeout. Cleanup of the latter recorded eight B70 engine resets and 61 unsuccessful responses before all four cards re-enumerated idle. The cell is nondeterministically unstable and receives no speed, curve, quality, deployment, or headline credit; unchanged retries and 24K/32K serving remain blocked pending a material runtime treatment.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-16k-a3-a4-runtime-instability.json" }, "0:24576": { "state": "missing", "label": "not yet tested" }, "0:32768": { "state": "missing", "label": "not yet tested" }, "1:16384": { "state": "missing", "label": "not yet tested" }, "1:24576": { "state": "missing", "label": "not yet tested" }, "1:32768": { "state": "missing", "label": "not yet tested" }, "2:16384": { "state": "quarantined", "label": "exact 16K stopped at 5,440 computed prompt tokens after scheduler-32 treatment; no output; four-card teardown events; diagnostic only", "reason": "The current-source MTP2 identity admitted exactly 40 blocks / 20,014 cache tokens in both bounded arms. Scheduler-64 stopped at 3,200 computed prompt tokens; the scheduler-32 arm was observed 2,240 tokens / 70% farther at 5,440, but the sole p16384/o128 treatment request still returned no output before the unchanged 300-second runtime response deadline. This is treatment evidence, not repeat-confirmed causality. Its shutdown was followed by eight card reset records and 58 unsuccessful card responses, so the frozen postflight rule fails and the treatment tranche is exhausted. No capability, speed, quality, parity, deployment, or headline credit is granted; current checks found no listener or owned residue and all cards are below 43 MiB.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-16512-scheduler32-attempt2-runtime-timeout-quarantine.json" }, "2:24576": { "state": "missing", "label": "not yet tested" }, "2:32768": { "state": "missing", "label": "not yet tested" }, "3:16384": { "state": "missing", "label": "not yet tested" }, "3:24576": { "state": "missing", "label": "not yet tested" }, "3:32768": { "state": "missing", "label": "not yet tested" }, "4:16384": { "state": "missing", "label": "not yet tested" }, "4:24576": { "state": "missing", "label": "not yet tested" }, "4:32768": { "state": "missing", "label": "not yet tested" } } }, { "id": "qwen-flash-next-tp-fit", "label": "Card-fit summary", "fixed": "Official FP8 child artifact, vLLM 658965050 plus kernels 2f829747, eager MTP0 text, zero prior context, automatic KV, and TP=EP. TP1 and TP2 need separate fit/offload designs; they are not claimed unsupported.", "row_axis": { "key": "mtp", "label": "Generation mode", "prefix": "MTP" }, "column_axis": { "key": "tp", "label": "TP", "prefix": "TP" }, "fixed_selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "runtime": "vLLM XPU 658965050 + kernels 2f829747", "runtime_family": "vLLM XPU", "parallel_profile": "TP=EP", "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "rows": [ 0 ], "columns": [ 1, 2, 4 ], "cells": { "0:1": { "state": "missing", "label": "separate fit/offload design required", "selectors": { "ep": 1 } }, "0:2": { "state": "missing", "label": "separate fit/offload design required", "selectors": { "ep": 2 } }, "0:4": { "state": "lab-screened", "label": "5.222 tok/s; quality caveat", "evidence_id": "qwen38-flash-next-fp8-tp4-attempt19", "packet_id": "qwen38-flash-next-fp8-tp4-research", "selectors": { "ep": 4 } } } }, { "id": "qwen-flash-next-graph-by-modality", "label": "Graph and modality summary", "fixed": "Official FP8 child artifact, vLLM 1372c62d plus staged kernels 2f829747, TP4/EP4, MTP0, zero prior context, and automatic KV. Eager text is screened; PIECEWISE text has bounded compile-resource evidence but no API, quality, or speed result; vision remains unmeasured.", "row_axis": { "key": "graph_mode", "label": "Graph mode", "prefix": "", "value_labels": { "off": "eager", "PIECEWISE": "graph" } }, "column_axis": { "key": "modality", "label": "Modality", "prefix": "", "value_labels": { "text": "text", "vision": "vision" } }, "fixed_selectors": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "tp": 4, "ep": 4, "parallel_profile": "TP=EP", "mtp": 0, "active_context_tokens": 0, "kv": "auto" }, "rows": [ "off", "PIECEWISE" ], "columns": [ "text", "vision" ], "cells": { "off:text": { "state": "lab-screened", "label": "5.224 tok/s; current-runtime Grade-C screen", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp0-current-a4", "packet_id": "qwen38-flash-next-fp8-tp4-mtp0-current-research" }, "off:vision": { "state": "missing", "label": "vision serving not measured" }, "PIECEWISE:text": { "state": "quarantined", "label": "all 131 shards loaded on four ranks; graph compile began; 30-GiB host-memory guard tripped; no API, quality, or speed", "reason": "Attempt 7 loaded the complete checkpoint on all four ranks and entered PIECEWISE torch.compile with one compile thread. The phase-aware guard latched below its 30-GiB host-memory floor before the later TTM/global-OOM teardown window. Graph capture did not complete, the API never became healthy, and no client, output, quality, replay, or speed row exists. This is Grade-D bounded resource evidence only.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-piecewise-graph-attempt7-result.json" }, "PIECEWISE:vision": { "state": "missing", "label": "graph vision serving not measured" } } } ], "coverage_contracts": [ { "id": "flash-next-text-serving-space", "label": "Text serving coverage", "description": "Exact text-serving combinations across topology, MTP depth, active context, and graph mode on the latest retained runtime. Older-runtime evidence is preserved in a separate archival contract and never silently transferred into these cells.", "fixed_selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "parallel_profile": "TP=EP" }, "axes": [ { "key": "revision", "label": "Weight revision", "values": [ "qwen38-flash-next" ] }, { "key": "artifact_id", "label": "Artifact", "values": [ "qwen38-flash-next-fp8-bcd9f01" ] }, { "key": "tp", "label": "Tensor parallel", "values": [ 1, 2, 4 ] }, { "key": "mtp", "label": "MTP depth", "values": [ 0, 1, 2, 3, 4 ] }, { "key": "active_context_tokens", "label": "Active context", "values": [ 0, 1024, 2048, 4096, 8192, 16384, 24576, 32768 ] }, { "key": "graph_mode", "label": "Graph mode", "values": [ "off", "PIECEWISE" ] }, { "key": "kv", "label": "KV cache", "values": [ "auto" ] }, { "key": "modality", "label": "Modality", "values": [ "text" ] } ], "rules": [ { "id": "default-gap", "match": { "revision": "*", "artifact_id": "*", "tp": "*", "mtp": "*", "active_context_tokens": "*", "graph_mode": "*", "kv": "*", "modality": "*" }, "state": "missing", "label": "not measured", "parent": "flash-next-qualification-backlog", "retry": { "status": "queued" } }, { "id": "tp4-mtp0-active0-piecewise-compile-resource-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 0, "active_context_tokens": 0, "graph_mode": "PIECEWISE", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "all shards loaded and graph compile began · 30-GiB host-memory guard tripped · no API, quality, or speed", "reason": "Attempt 7 loaded all 131 shards on all four ranks and entered PIECEWISE graph compilation with one compile thread. The phase-aware host-memory guard latched below 30 GiB before the later TTM/global-OOM teardown window. Graph capture did not complete; no healthy API, client, output, quality result, replay result, or speed row exists. This is Grade-D bounded resource evidence only.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-current-piecewise-graph-attempt7-result.json", "parent": "flash-next-piecewise-host-memory-design", "retry": { "status": "blocked-on-material-memory-design", "trigger": "a materially different preregistered compile-memory treatment after clean host recovery" } }, { "id": "tp1-mtp0-active0-static-fit-closure", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 1, "mtp": 0, "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "closed", "label": "fit closed on this host · exact 169.41-GiB text target", "reason": "Header accounting across all 131 shards leaves 169.413485 GiB for the text-only MTP0 target after excluding 0.836199 GiB of vision weights and 2.512733 GiB of MTP weights. One 31.890625-GiB B70 plus this host's 125.652466 GiB physical RAM is 11.870394 GiB short before runtime, cache, workspace, or operating-system headroom. This is a speedless Grade-D fit boundary, not a boot or performance estimate.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp1-ep1-eager-mtp0-active0-static-fit-boundary.json", "parent": "flash-next-tp1-memory-design", "retry": { "status": "blocked-on-material-memory-design", "trigger": "at least 192 GiB host RAM with qualified expert offload, or a separately qualified compression or streaming design" } }, { "id": "context16k-anchor-required", "match": { "revision": "*", "artifact_id": "*", "tp": "*", "mtp": "*", "active_context_tokens": 16384, "graph_mode": "*", "kv": "*", "modality": "*" }, "state": "missing", "label": "16K capability and quality anchor required", "reason": "No unquarantined 16K cell exists. A completed request alone is insufficient: qualification needs exact depth, semantic quality, repeat stability, controlled teardown, and a clean exact-topology postflight under the exact cell identity.", "parent": "flash-next-context16k-qualification-backlog", "retry": { "status": "needs-qualified-16k-anchor", "trigger": "one preregistered MTP0 or MTP1 anchor after clean exact-topology recovery" } }, { "id": "context24k-anchor-required", "match": { "revision": "*", "artifact_id": "*", "tp": "*", "mtp": "*", "active_context_tokens": 24576, "graph_mode": "*", "kv": "*", "modality": "*" }, "state": "missing", "label": "24K capability and quality anchor required", "reason": "No 24K Flash-Next request has been qualified. This cell needs a measured capacity identity and exact-depth capability receipt before quality, speed, or any estimate can receive credit.", "parent": "flash-next-context24k-qualification-backlog", "retry": { "status": "needs-measured-24k-anchor", "trigger": "qualify TP4 eager MTP0 first, then transfer only explicitly justified bounds" } }, { "id": "context32k-anchor-required", "match": { "revision": "*", "artifact_id": "*", "tp": "*", "mtp": "*", "active_context_tokens": 32768, "graph_mode": "*", "kv": "*", "modality": "*" }, "state": "missing", "label": "32K capability and quality anchor required", "reason": "No 32K Flash-Next request has been qualified. This cell needs a measured capacity identity and exact-depth capability receipt before quality, speed, or any estimate can receive credit.", "parent": "flash-next-context32k-qualification-backlog", "retry": { "status": "needs-measured-32k-anchor", "trigger": "qualify TP4 eager MTP0 after the 24K anchor, then transfer only explicitly justified bounds" } }, { "id": "mtp2-context24k-treatment-blocked", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 2, "active_context_tokens": 24576, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP2 24K blocked by exhausted 16K treatment", "reason": "Both bounded 16K MTP2 arms produced no output and the scheduler-32 arm failed postflight. A deeper launch would not answer a new question and is blocked until a material runtime treatment qualifies at 16K.", "parent": "flash-next-mtp2-deep-context-runtime-block", "retry": { "status": "blocked-on-material-runtime-treatment", "trigger": "new runtime fix that completes and cleanly tears down the 16K MTP2 gate" } }, { "id": "mtp2-context32k-treatment-blocked", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 2, "active_context_tokens": 32768, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP2 32K blocked by exhausted 16K treatment", "reason": "Both bounded 16K MTP2 arms produced no output and the scheduler-32 arm failed postflight. A deeper launch would not answer a new question and is blocked until a material runtime treatment qualifies at 16K.", "parent": "flash-next-mtp2-deep-context-runtime-block", "retry": { "status": "blocked-on-material-runtime-treatment", "trigger": "new runtime fix that completes and cleanly tears down the 16K MTP2 gate" } }, { "id": "mtp3-context16k-runtime-treatment-required", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 3, "active_context_tokens": 16384, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP3 16K blocked pending shallower stability treatment", "reason": "MTP3 did not complete its bounded 8K request. A 16K launch is not authorized until a material runtime treatment completes the exact 8K gate with clean postflight.", "parent": "flash-next-mtp3-deep-context-runtime-block", "retry": { "status": "blocked-on-qualified-8k-gate", "trigger": "material treatment plus completed exact-8K MTP3 receipt and clean postflight" } }, { "id": "mtp3-context24k-runtime-treatment-required", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 3, "active_context_tokens": 24576, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP3 24K blocked pending shallower stability treatment", "reason": "MTP3 did not complete its bounded 8K request. Deeper launches are blocked until a material treatment qualifies 8K and 16K sequentially.", "parent": "flash-next-mtp3-deep-context-runtime-block", "retry": { "status": "blocked-on-qualified-shallower-gates", "trigger": "qualified exact-8K and exact-16K MTP3 receipts with clean postflight" } }, { "id": "mtp3-context32k-runtime-treatment-required", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 3, "active_context_tokens": 32768, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP3 32K blocked pending shallower stability treatment", "reason": "MTP3 did not complete its bounded 8K request. Deeper launches are blocked until a material treatment qualifies 8K, 16K, and 24K sequentially.", "parent": "flash-next-mtp3-deep-context-runtime-block", "retry": { "status": "blocked-on-qualified-shallower-gates", "trigger": "qualified exact-8K, exact-16K, and exact-24K MTP3 receipts with clean postflight" } }, { "id": "mtp4-context16k-parity-treatment-required", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 4, "active_context_tokens": 16384, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP4 16K blocked pending parity treatment", "reason": "MTP4 completed 8K only as diagnostic evidence because it diverged from the frozen authority. Deeper capability work is blocked until a same-runtime authority or material parity treatment resolves that shallower cell.", "parent": "flash-next-mtp4-deep-context-parity-block", "retry": { "status": "blocked-on-parity-treatment", "trigger": "same-runtime authority design or material parity treatment that qualifies exact 8K" } }, { "id": "mtp4-context24k-parity-treatment-required", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 4, "active_context_tokens": 24576, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP4 24K blocked pending parity treatment", "reason": "MTP4 completed 8K only as diagnostic evidence because it diverged from the frozen authority. Deeper work is blocked until parity is resolved and 16K qualifies cleanly.", "parent": "flash-next-mtp4-deep-context-parity-block", "retry": { "status": "blocked-on-qualified-shallower-gates", "trigger": "qualified exact-8K and exact-16K MTP4 receipts under a same-runtime authority" } }, { "id": "mtp4-context32k-parity-treatment-required", "match": { "revision": "*", "artifact_id": "*", "tp": 4, "mtp": 4, "active_context_tokens": 32768, "graph_mode": "off", "kv": "*", "modality": "*" }, "state": "missing", "label": "MTP4 32K blocked pending parity treatment", "reason": "MTP4 completed 8K only as diagnostic evidence because it diverged from the frozen authority. Deeper work is blocked until parity is resolved and all shallower depth gates qualify cleanly.", "parent": "flash-next-mtp4-deep-context-parity-block", "retry": { "status": "blocked-on-qualified-shallower-gates", "trigger": "qualified exact-8K, exact-16K, and exact-24K MTP4 receipts under a same-runtime authority" } }, { "id": "mtp1-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 1, "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "9.372 tok/s matched MTP1 research screen", "reason": "All 26 MTP0 baseline comparisons matched with thinking disabled on both clients; fixed-set repeats passed 16/16, the small cache-zero needle passed, and all three speed rows returned the target hash. Inherited short quality remains 5/7. The separate exact-4K headroom32 screen does not overwrite this configured-512 cell.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp1-a3", "packet_id": "qwen38-flash-next-fp8-tp4-mtp1-research", "parent": "flash-next-mtp1-a3", "retry": { "status": "needs-full-quality" } }, { "id": "mtp2-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 2, "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "11.895 tok/s variable matched MTP2 research screen", "reason": "All 26 bounded MTP0 comparisons matched with thinking disabled on both clients; fixed-set repeats passed 16/16, the small cache-zero needle passed, and all three speed rows returned the target hash. Inherited short quality remains 5/7, and the rows span 29.61% of their median. The separate active-4K selector is quarantined and does not lower this configured-512 cell.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp2-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp2-research", "parent": "flash-next-mtp2-a1", "retry": { "status": "needs-speed-stability-and-full-quality" } }, { "id": "mtp3-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 3, "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "14.889 tok/s variable matched MTP3 research screen", "reason": "All 26 bounded MTP0 comparisons matched with thinking disabled on both clients; fixed-set repeats passed 16/16, the small cache-zero needle passed, and all three speed rows returned the target hash. The needle was only 317 actual prompt tokens, inherited short quality remains 5/7, and the speed rows declined monotonically across a wide range.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp3-a4", "packet_id": "qwen38-flash-next-fp8-tp4-mtp3-research", "parent": "flash-next-mtp3-a4", "retry": { "status": "needs-full-quality" } }, { "id": "mtp4-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 4, "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "20.727 tok/s matched MTP4 research screen", "reason": "All 26 bounded MTP0 comparisons matched with thinking disabled on both clients; fixed-set repeats passed 16/16, the small cache-zero needle passed, all three corrected speed rows returned the target hash, and cumulative acceptance was 1,716/1,716. Inherited short quality remains 5/7; the separate exact-4K MTP4 selector is quarantined and does not lower this cell.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp4-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp4-research", "parent": "flash-next-mtp4-a1", "retry": { "status": "needs-full-quality" } }, { "id": "mtp1-context1k-first-request-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 1, "active_context_tokens": 1024, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "first 1K request stopped at 768 computed tokens · no output", "reason": "The exact source/runtime identity booted from the local-NVMe model, passed four-rank preflight, reported all four 12.22-GiB placements, admitted 32 cache blocks and became healthy. Its first 1K request reached the unchanged 300-second worker-response gate during sampling with 768 computed prompt tokens and zero output. Request two and the separate 2K boot were skipped under the preregistered stop rule. Existing MTP1 512 and 4K passes are unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-1536-context-attempt1-bounded-negative.json", "parent": "flash-next-mtp1-context1k-a1", "retry": { "status": "blocked-on-runtime-fix", "trigger": "material first-request completion fix plus fresh four-rank boot" } }, { "id": "mtp1-context2k-client-and-engine-timeout-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 1, "active_context_tokens": 2048, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "client timeout after 360 s · 448 computed tokens · no output", "reason": "The exact local-NVMe source/runtime identity passed four-rank, placement, 32-block cache, capacity, identity, and health gates. Request one had a zero-byte completion body and no output token recorded when the fixed client bound expired. The subsequent engine diagnostic reported 448 computed prompt tokens and zero output, after which the engine independently reported its own sampling timeout. Request two was blocked. The post-failure teardown window recorded compute- and copy-class resets on all four cards; all devices were rediscovered, but no post-reset collective was run. Existing MTP1 configured-512 and exact-4K passes are unchanged.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-3072-context-attempt2-bounded-negative.json", "parent": "flash-next-mtp1-context2k-a2", "retry": { "status": "blocked-on-runtime-fix", "trigger": "material first-request completion fix plus fresh post-reset four-rank validation" } }, { "id": "mtp1-context4k-headroom32-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 1, "active_context_tokens": 4096, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "8.904 tok/s decode · 232.1 s TTFT · headroom32", "reason": "The 32-block recipe passed all 26 sealed MTP0 comparisons, 16/16 repeats, the exact cache-zero 4096-token needle, a formal p4096/o128 gate, and all three p4096/o256 rows with one accepted-target hash. Median wall output was 0.981 tok/s and cumulative draft acceptance was 528/539. This is a Grade-C support cell, not a minimum-cache claim; MTP3 remains the preferred exact-4K recipe.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp1-context4k-headroom32-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp1-research", "parent": "flash-next-mtp1-context4k-headroom32-a1", "retry": { "status": "needs-clean-boot-full-quality" } }, { "id": "mtp1-context8k-cross-runtime-parity-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 1, "active_context_tokens": 8192, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "exact 8K completed · cross-runtime parity mismatch at token 73 · diagnostic only", "reason": "The exact current-source identity passed fresh four-rank, placement, 32-block cache, capacity, served-model, and health gates and exposed 13,516 cache tokens. Its sole p8192/o128 request completed all 25 generic gates with exact usage, zero cache reuse, and positive MTP1 counters, but diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 72. The 4.151 tok/s rate and 953.3-second TTFT receive no speed or quality credit. Owned shutdown passed its cleanup gates with all four cards idle and no B70-addressed event; the intentional stop retained the known shutdown-time output-handler notice and one shared-memory cleanup warning. Seven corrected local-NVMe records separately block clean-host wording.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp1-8448-context-attempt1-parity-quarantine.json", "parent": "flash-next-mtp1-context8k-a1", "retry": { "status": "blocked-on-parity-treatment", "trigger": "material parity treatment, same-runtime authority design, or a new preregistration with a distinct causal question" } }, { "id": "mtp2-context1k-host-health-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 2, "active_context_tokens": 1024, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "two exact target-parity requests passed; clean-host gate failed on corrected NVMe events", "reason": "The exact local-NVMe source/runtime identity passed fresh four-rank, placement, 32-block cache, capacity, identity, and health gates. Both authorized requests returned exactly 1,024 prompt and 256 output tokens, the frozen MTP0 text hash, zero cache reuse, identical text, and perfect MTP2 acceptance at both positions. Request one observed 10.683 tok/s after first text and the repeat sentinel 12.642 tok/s, but 11 corrected events for local NVMe 0000:01:00.0 appeared after the frozen journal cutoff. The strict clean-host rule quarantines the cell and gives both rates diagnostic-only status; no B70 event appeared.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-1536-context-attempt1-host-quarantine.json", "parent": "flash-next-mtp2-context1k-a1", "retry": { "status": "blocked-on-clean-storage-window", "trigger": "clean local-NVMe link or identical verified model on storage with a clean post-cutoff host window" } }, { "id": "mtp2-context2k-target-parity-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 2, "active_context_tokens": 2048, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "completed exact-2K request · target-parity mismatch at token 13 · no speed credit", "reason": "The local-NVMe boot passed four-rank, placement, 32-block cache, capacity, identity, and health gates. Request one completed with 2,048 prompt and 128 output tokens, zero cache reuse, and positive MTP2 counters, but diverged from the frozen MTP0 authority at zero-based generated-token index 12. Request two was blocked. This is a scoped cross-lane parity mismatch, not isolated proof that MTP2 caused the divergence.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-3072-context-attempt1-parity-quarantine.json", "parent": "flash-next-mtp2-context2k-a1", "retry": { "status": "blocked-on-parity-fix", "trigger": "material target-parity treatment plus a new preregistration" } }, { "id": "mtp2-context4k-headroom32-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 2, "active_context_tokens": 4096, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "9.893 tok/s decode · 263.3 s TTFT · headroom32", "reason": "The 32-block recipe passed all 26 sealed MTP0 comparisons, 16/16 repeats, the exact cache-zero 4096-token needle, a formal p4096/o128 gate, and all three p4096/o256 rows with one accepted-target hash. Median wall output was 0.891 tok/s and cumulative draft acceptance was 719/748. The failed 21-block arm remains linked as history; this pass does not prove a speed gain, causal cache mechanism, or minimum cache. MTP3 remains preferred.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp2-context4k-headroom32-a2", "packet_id": "qwen38-flash-next-fp8-tp4-mtp2-research", "parent": "flash-next-mtp2-context4k-headroom32-a2", "retry": { "status": "needs-clean-boot-full-quality" } }, { "id": "mtp2-context8k-target-parity-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 2, "active_context_tokens": 8192, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "completed exact-8K request · parity mismatch at token 27 · no speed credit", "reason": "The exact 32-block current-source identity exposed 11,264 cache tokens and completed one p8192/o128 request with exact 8192/128/8320 usage, zero cache reuse, 128 token IDs, and positive MTP2 counters at positions zero and one. Its token array first diverged from the frozen MTP0 authority at zero-based generated-token index 26. The observed 6.235 tok/s and 649.7-second TTFT are diagnostic only. The authority used a different vLLM commit and cache allocation, so this is a scoped cross-runtime parity quarantine rather than isolated proof that MTP2 caused the difference. Teardown passed with no B70 event; corrected storage/root-port events block clean-host wording.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-8448-context-attempt2-parity-quarantine.json", "parent": "flash-next-mtp2-context8k-a2", "retry": { "status": "blocked-on-parity-fix", "trigger": "material parity treatment plus a new preregistration" } }, { "id": "mtp3-context1k-external-signal-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 3, "active_context_tokens": 1024, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "external SIGTERM during request one · six accepted draft tokens · no completed response or speed", "reason": "The exact 25-block local-NVMe identity passed source, runtime, fresh four-rank, placement, cache, capacity, served-identity, and health gates. Request one began under the frozen protocol, then the server received an external SIGTERM before the response completed. No request JSON, usage, output hash, or performance result exists. Partial server metrics showed six drafted and six accepted tokens with 1.000 acceptance at all three MTP3 positions, but those counters receive no parity, speed, quality, or deployment credit. Request two was not sent.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-1536-context-attempt1-external-stop.json", "parent": "flash-next-mtp3-context1k-a1", "retry": { "status": "blocked-on-new-preregistration", "trigger": "new preregistration that explicitly addresses external process-session continuity" } }, { "id": "mtp3-context2k-target-parity-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 3, "active_context_tokens": 2048, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "completed exact-2K request · target-parity mismatch at token five · no speed credit", "reason": "The exact source/runtime identity booted from local NVMe, passed four-rank preflight, reported all four 12.22-GiB placements, admitted the 25-block cache and became healthy. Request one completed the generic exact-depth gate with 2,048 prompt and 128 output tokens, zero cache reuse, and positive MTP3 counters, but its token array diverged from the frozen MTP0 authority at zero-based generated-token index 4. The preregistered rule blocked request two. This is a scoped cross-lane parity mismatch, not isolated proof that MTP3 caused the divergence or a universal semantic-quality failure.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-3072-context-attempt1-parity-quarantine.json", "parent": "flash-next-mtp3-context2k-a1", "retry": { "status": "blocked-on-parity-fix", "trigger": "material target-parity treatment plus a new preregistration" } }, { "id": "mtp3-context4k-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 3, "active_context_tokens": 4096, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "15.502 tok/s decode · 187.9 s TTFT exact-4K", "reason": "All 26 sealed MTP0 4K comparisons matched, fixed-set repeats passed 16/16, and the exact cache-zero 4096-token needle plus formal gate passed. The three service rows returned one accepted-target hash. Long-prompt prefill dominates: median wall output was 1.246 tok/s and median TTFT was 187.9 seconds; inherited short quality remains 5/7. The official-thinking transfer completed 19 of 25 required rows; every completed row passed and matched MTP0, but the repeated-session boundary stopped request 20 and left six rows unrun. The outcome is inconclusive and unqualified, not an answer-quality failure.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp3-context4k-a1", "packet_id": "qwen38-flash-next-fp8-tp4-mtp3-research", "parent": "flash-next-mtp3-context4k-a1", "retry": { "status": "blocked-on-repeated-session-stability", "trigger": "separately-preregistered bounded stability gate" } }, { "id": "mtp3-context8k-client-timeout-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 3, "active_context_tokens": 8192, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "900-second client timeout · no completed receipt or speed", "reason": "The exact current-source identity passed fresh four-rank, placement, 32-block cache, capacity, served-model, and health gates and exposed 9,654 cache tokens. Its sole p8192/o128 request entered inference but reached the fixed 900-second client bound without a completed response receipt or any durably recorded output token. No usage, MTP counter, parity, rate, or TTFT result exists. The owned failed-request shutdown path passed postflight with all four cards idle and no B70-addressed event; corrected local-NVMe records separately block clean-host wording.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp3-8448-context-attempt1-bounded-negative.json", "parent": "flash-next-mtp3-context8k-a1", "retry": { "status": "blocked-on-runtime-treatment", "trigger": "material completion treatment or separately justified progress evidence, plus new preregistration and fresh four-rank preflight; a longer bound alone is insufficient" } }, { "id": "mtp4-context1k-teardown-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 4, "active_context_tokens": 1024, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "two exact target-parity requests passed · detached teardown gate failed", "reason": "The exact 29-block local-NVMe identity passed every startup gate. Both authorized requests returned 1,024 prompt and 256 output tokens, the frozen MTP0 text hash, zero cache reuse, identical text, and perfect MTP4 acceptance at positions zero through three. The observed 13.326 and 17.291 tok/s rates are diagnostic only because the exact stop sentinel ended the timeout supervisor without reaching the detached server group. Direct recovery produced an orderly shutdown, but the frozen teardown rule quarantines the cell. Seven corrected-only local-NVMe records independently block clean-host qualification; no B70 event appeared.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-1536-context-attempt1-teardown-quarantine.json", "parent": "flash-next-mtp4-context1k-a1", "retry": { "status": "blocked-on-lifecycle-and-clean-storage", "trigger": "separately tested descendant-aware supervisor plus a clean storage-link window" } }, { "id": "mtp4-context2k-no-output-reset-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 4, "active_context_tokens": 2048, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "client and engine timeouts · 384 computed tokens · no output · four-card teardown resets", "reason": "The exact 29-block configured-3072 identity passed every startup gate. Request one reached the fixed 360-second client bound without a response receipt; about five seconds later the engine independently reported its own sampling timeout at 384 computed prompt tokens and zero output. No HTTP status, usage, parity hash, MTP counter delta, or speed exists, and request two was blocked. The corrected lifecycle controller returned zero and left no listener, recorded process, compile path, or RPC path, but the teardown window recorded compute- and copy-class resets on all four B70s. All cards were rediscovered at low memory use; no post-reset collective or known-good generation canary was run.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-3072-context-attempt1-bounded-negative.json", "parent": "flash-next-mtp4-context2k-a1", "retry": { "status": "blocked-on-runtime-fix-and-post-reset-qualification", "trigger": "material first-request completion treatment, new preregistration, and full post-reset recovery qualification" } }, { "id": "mtp4-context4k-worker-timeout-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 4, "active_context_tokens": 4096, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "worker timeout at 3,904 computed tokens · no speed", "reason": "The exact 29-block configuration became healthy and admitted 4,352 tokens, but the 4K quality request exceeded the worker-response deadline during token sampling and the engine stopped. Cleanup was followed by card-engine resets on all four B70 addresses. No durable quality JSON or timing result was produced.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260827-tp4-mtp4-4352-attempt1-bounded-negative.json", "parent": "flash-next-mtp4-context4k-a1", "retry": { "status": "blocked-on-runtime-fix", "trigger": "material worker-stall fix plus post-reset four-rank preflight" } }, { "id": "mtp4-context8k-cross-runtime-parity-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 4, "active_context_tokens": 8192, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "exact 8K completed · MTP4 mechanism passed · parity mismatch at token 27 · diagnostic only", "reason": "The exact 36-block current-source identity completed one p8192/o128 request with exact 8192/128/8320 usage, zero cache reuse, all 25 generic depth gates, and positive MTP4 accepted-token deltas [30,24,18,14] whose sum equals 86. Its output first differs from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 26. The 4.026 tok/s rate and 918.4-second TTFT are diagnostic only and do not isolate MTP4 as the cause. The owned completed-classification path passed postflight and returned all four cards idle; shutdown warnings and six corrected local-NVMe endpoint records remain disclosed, with no B70-addressed event.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp4-8448-context-attempt2-parity-quarantine.json", "parent": "flash-next-mtp4-context8k-a2", "retry": { "status": "blocked-on-parity-treatment-or-same-runtime-authority", "trigger": "material parity treatment, a same-runtime authority design, or a separately preregistered distinct causal question" } }, { "id": "mtp0-current-short-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 0, "active_context_tokens": 0, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "5.224 tok/s · current-runtime short Grade-C screen", "reason": "The exact current-runtime TP4/EP4 MTP0 identity passed fresh four-rank preflight, an exact cache-zero recovery canary, 6/7 semantic quality with only the known code-expression miss, 16/16 repeats, and card-clean teardown. Three established p146/o256/c1 after-first-text rows measured 5.316, 5.224, and 5.219 tok/s. The short harness does not retain per-row cache or finish fields, and this is not the fixed realistic final suite, a clean-host replay, or deployment qualification.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp0-current-a4", "packet_id": "qwen38-flash-next-fp8-tp4-mtp0-current-research", "parent": "flash-next-mtp0-current-short-a4", "retry": { "status": "needs-fresh-server-clean-host-final-suite" } }, { "id": "mtp0-current-context4k-text-screen", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 0, "active_context_tokens": 4096, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "lab-screened", "label": "4.758 tok/s decode · 147.5 s TTFT · current runtime exact-4K", "reason": "After the current-runtime 6/7 semantic, 16/16 repeat, and exact cache-zero 4K needle gates, two exact p4096/o128 requests both returned 4096/128/4224 usage, zero cache reuse, length stops, 128 token IDs, and one output-token hash. Their conventional 99-interval rates were 4.720 and 4.795 tok/s; the same-boot median is 4.758 tok/s with 147.5-second median TTFT. The output hash matches the retained legacy target authority, but fresh-server and clean-host replay remain open.", "evidence_id": "qwen38-flash-next-fp8-tp4-mtp0-current-context4k-a4", "packet_id": "qwen38-flash-next-fp8-tp4-mtp0-current-research", "parent": "flash-next-mtp0-current-context4k-a4", "retry": { "status": "needs-fresh-server-clean-host-repeat" } }, { "id": "mtp0-context16k-current-source-generic-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 0, "active_context_tokens": 16384, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "16K generic request completed once · semantic repeat corrupted · separate fresh server timed out · diagnostic only", "reason": "The current-source 33-block identity completed one generic p16384/o128 request at a diagnostic-only 5.219 tok/s. A semantic successor then completed its first 16,213-token request correctly, returned corrupted repeated text on the identical same-server request despite zero reported cache use, and a separately started fresh server stopped at 1,600 computed prompt tokens before first output with an RPC sampling timeout. Cleanup of the fresh-server failure recorded eight B70 engine resets and 61 unsuccessful responses before all four cards re-enumerated idle. This is nondeterministic long-context runtime-stability evidence, not a speed or quality result.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp0-16k-a3-a4-runtime-instability.json", "parent": "flash-next-mtp0-context16k-current-source-a2", "retry": { "status": "blocked-on-material-runtime-treatment-and-fresh-four-card-health-gate" } }, { "id": "mtp2-context16k-current-source-scheduler-treatment-exhausted-quarantine", "match": { "revision": "qwen38-flash-next", "artifact_id": "qwen38-flash-next-fp8-bcd9f01", "tp": 4, "mtp": 2, "active_context_tokens": 16384, "graph_mode": "off", "kv": "auto", "modality": "text" }, "state": "quarantined", "label": "MTP2 scheduler treatment stopped at 5,440/16,384 computed prompt tokens · no output · card-event postflight failed", "reason": "The current-source MTP2 identity admitted exactly 40 blocks / 20,014 cache tokens in both bounded arms. Scheduler-64 stopped at 3,200 computed prompt tokens; the scheduler-32 arm was observed 2,240 tokens / 70% farther at 5,440, but the sole p16384/o128 treatment request still returned no output before the unchanged 300-second runtime response deadline. This is treatment evidence, not repeat-confirmed causality. Its shutdown was followed by eight card reset records and 58 unsuccessful card responses, so the frozen postflight rule fails and the treatment tranche is exhausted. No capability, speed, quality, parity, deployment, or headline credit is granted; current checks found no listener or owned residue and all cards are below 43 MiB.", "evidence": "experiments/qwen38-flash-next-fp8-b70/data/20260828-tp4-mtp2-16512-scheduler32-attempt2-runtime-timeout-quarantine.json", "parent": "flash-next-mtp2-context16k-current-source-a2-scheduler32", "retry": { "status": "treatment-exhausted-no-further-mtp2-active16k-retry" } } ] }, { "id": "flash-next-fixed-vision-serving-space", "label": "Fixed-vision serving coverage", "description": "Vision capability and quality combinations under one versioned fixed-image fixture. Context depth is intentionally excluded until a vision anchor establishes meaningful image-plus-text token accounting.", "fixed_selectors": { "runtime": "vLLM XPU 1372c62d + staged kernels 2f829747", "runtime_family": "vLLM XPU", "parallel_profile": "TP=EP", "active_context_tokens": 0, "modality": "vision", "workload_profile": "fixed-vision-fixture-v1-required" }, "axes": [ { "key": "revision", "label": "Weight revision", "values": [ "qwen38-flash-next" ] }, { "key": "artifact_id", "label": "Artifact", "values": [ "qwen38-flash-next-fp8-bcd9f01" ] }, { "key": "tp", "label": "Tensor parallel", "values": [ 1, 2, 4 ] }, { "key": "mtp", "label": "MTP depth", "values": [ 0, 1, 2, 3, 4 ] }, { "key": "graph_mode", "label": "Graph mode", "values": [ "off", "PIECEWISE" ] }, { "key": "kv", "label": "KV cache", "values": [ "auto" ] } ], "rules": [ { "id": "fixed-vision-anchor-required", "match": { "revision": "*", "artifact_id": "*", "tp": "*", "mtp": "*", "graph_mode": "*", "kv": "*" }, "state": "missing", "label": "fixed-image capability and quality anchor required", "reason": "No versioned fixed-image Flash-Next workload has been run on this runtime. Define one fixture and qualify TP4 eager MTP0 before measuring or estimating vision topology, graph, or MTP variants.", "parent": "flash-next-fixed-vision-qualification-backlog", "retry": { "status": "needs-fixed-vision-anchor", "trigger": "versioned fixed-image fixture plus TP4 eager MTP0 capability and quality receipt" } } ] } ] }