{ "format": "neural-download-model-family-v1", "id": "laguna-s", "primary_packet_id": "laguna-s-2.1-int4-b70-125tps-20260731", "name": "Laguna S", "display_name": "Laguna S · 2.1", "publisher": "Poolside", "updated_at": "2026-08-23", "summary": "Poolside's Laguna S 2.1, a coding assistant model. Runs across four Arc Pro B70 cards with a draft model speeding up generation; measured as a coding-first deployment.", "architecture": { "target_class": "LagunaForCausalLM", "draft_class": "DFlashLagunaForCausalLM", "deployment_modalities": ["text"], "evidence": "repro/laguna-s-2.1-int4-b70-102tps-20260726/evidence/record-run/server.log" }, "weight_revisions": [ { "id": "laguna-s-2.1-int4-4bbfc28", "label": "Laguna S 2.1 INT4", "role": "target/verifier checkpoint", "repository": "poolside/Laguna-S-2.1-INT4", "revision": "4bbfc285f2f8b3b6b526274c133b7b17aae6c8cb", "model_manifest": "repro/laguna-s-2.1-int4-b70-102tps-20260726/manifests/model-release-files.sha256" } ], "auxiliary_artifacts": [ { "id": "laguna-s-2.1-dflash-int4-5e07c24", "label": "Laguna S 2.1 DFlash INT4", "role": "speculative draft checkpoint; not a family weight revision or target model", "repository": "poolside/Laguna-S-2.1-DFlash-INT4", "revision": "5e07c246915c86dc6920fead03d019989224f2ba", "model_manifest": "repro/laguna-s-2.1-int4-b70-102tps-20260726/manifests/model-release-files.sha256" } ], "transfer_scope": { "status": "single pinned target/draft pair with revision-specific runtime and native artifacts", "transfers": [ "The exact target/draft pairing, width-12 verifier contract, DFlash depth, and fail-closed quality methodology transfer only when their pinned identities are preserved.", "The current 125.46 tok/s lane is an evolution of the earlier sealed four-card packet, not a replacement for its independently pinned runtime provenance." ], "does_not_transfer": [ "performance to other Laguna checkpoints or quantizations", "performance to other GPU counts, TP/EP layouts, or KV dtypes", "the draft FP8 W8A16 projection treatment to the INT4 target label", "the historical 100-event submission convention to conventional interval throughput", "the originating-host result to a portable or non-originating-host replay" ], "evidence": "repro/laguna-s-2.1-int4-b70-125tps-20260731/README.md" }, "model_signals": { "b70_fit": { "band": "high", "scope": "four-card text deployment on Intel Arc Pro B70", "basis": "The pinned target/draft pair has exact-output, cache-zero, one-start records on TP4+EP4.", "reviewed_at": "2026-08-23" }, "quality_evidence": { "band": "strong-deployment-evidence", "scope": "sealed 13-prompt teacher exactness, cache-zero, graph/treatment markers, and idle-host gates; not a general intelligence score", "evidence": [ "data/laguna-s-2.1-width12-dflash-fp8-record-20260726.json", "data/laguna-shared-elementwise-m12-record-20260731.json" ] }, "popularity": { "state": "not-scored", "reason": "No dated family-level popularity snapshot is stored for the exact target/draft pair." } }, "dimensions": { "weight_revision": ["laguna-s-2.1-int4-4bbfc28"], "target_quantization": ["INT4 compressed-tensors"], "draft_quantization": ["INT4 checkpoint with runtime E4M3FN W8A16 draft projections"], "runtime": ["vLLM XPU"], "cards": [4], "tp": [4], "ep": [4], "prompt_tokens": [1024, 4096, 8192, 16384, 24576, 32640], "configured_max_context_tokens": [32768], "speculative_method": ["DFlash"], "draft_depth": [11], "verifier_width": [12], "graph": ["PIECEWISE capture size 12"], "kv": ["bfloat16"] }, "packets": [ { "id": "laguna-s-2.1-int4-b70-125tps-20260731", "label": "Laguna S 2.1 INT4 · TP4+EP4 + DFlash11", "revision": "laguna-s-2.1-int4-4bbfc28", "draft_revision": "laguna-s-2.1-dflash-int4-5e07c24", "quantization": "INT4 target + INT4 DFlash draft / BF16 KV", "runtime": "vLLM XPU 1a7f61feffbc61b21b73f812d231c7426386ccdc", "cards": 4, "status": "candidate", "evidence_level": "B70-verified originating-host replay", "coverage": ["decode", "TTFT", "quality", "runtime provenance", "recipe"], "projection": {"model": "laguna_s_2.1", "quant": "AutoRound INT4", "runtime": "vllm", "spec": "dflash:11", "prompt_tokens": 128, "output_tokens": 128}, "manifest": "packages/laguna-s-2.1-int4-b70-125tps/package.json" }, { "id": "laguna-s-2.1-int4-b70-102tps-20260726", "label": "Laguna S 2.1 INT4 · historical TP4+EP4 + DFlash11", "revision": "laguna-s-2.1-int4-4bbfc28", "draft_revision": "laguna-s-2.1-dflash-int4-5e07c24", "quantization": "INT4 target + INT4 DFlash draft / BF16 KV", "runtime": "vLLM XPU e596ef1543466ae1a05e5bb8091f58872e2b18ba", "cards": 4, "status": "superseded-approved-record", "evidence_level": "B70-verified sealed history", "coverage": ["decode", "TTFT", "DFlash acceptance", "quality", "portable evidence audit", "recipe"], "manifest": "repro/laguna-s-2.1-int4-b70-102tps-20260726/README.md" }, { "id": "laguna-s-2.1-int4-b70-long-context-screen-20260802", "label": "Laguna S 2.1 INT4 · TP4+EP4 long-context screen", "revision": "laguna-s-2.1-int4-4bbfc28", "draft_revision": "laguna-s-2.1-dflash-int4-5e07c24", "quantization": "INT4 target + INT4 DFlash draft / BF16 KV", "runtime": "vLLM XPU 1a7f61feffbc61b21b73f812d231c7426386ccdc + XPU kernels 99886d783372e621941228250091dc8ebdc1595d", "cards": 4, "status": "bounded-research-screen-not-q1-exact", "evidence_level": "B70-screened retrieval evidence; not promotion-grade", "coverage": ["1K through 32,640 prompt tokens", "decode", "prefill", "TTFT", "DFlash acceptance", "retrieval", "q1 exactness audit"], "grades": { "evidence": { "grade": "C", "scope": "composed TP4+EP4 long-context retrieval screen from 1,024 through 32,640 prompt tokens", "basis": "exact prompt lengths, cache-zero, 128-token completions, retrieval, and post-32K sentinels passed, but the profile was composed from three runs and q1 output equality failed at every 4K-and-longer point", "reviewed_at": "2026-08-23", "evidence": ["data/laguna-s-2.1-xpu-b70/long-context-baseline-gpu080-swap24g-20260802.json"] } }, "manifest": "data/laguna-s-2.1-xpu-b70/long-context-baseline-gpu080-swap24g-20260802.json" } ], "run_measurements": [ { "id": "laguna-s-current-width12-dflash11-record", "state": "lab-measured", "revision": "laguna-s-2.1-int4-4bbfc28", "variant": "INT4 target + INT4 DFlash draft with runtime E4M3FN W8A16 draft projections", "runtime": "vLLM XPU 1a7f61feffbc61b21b73f812d231c7426386ccdc + XPU kernels 99886d783372e621941228250091dc8ebdc1595d", "config": {"record_order": 2, "cards": 4, "tp": 4, "ep": 4, "dflash": 11, "verifier_width": 12, "graph": "PIECEWISE", "capture_size": 12, "kv": "bfloat16"}, "workload": "sealed 13-prompt, one-start, cache-zero record gate with one active generation; conventional 99-interval metric", "metrics": { "decode_tok_s": [125.4619731637751], "ttft_ms": [5953.834546999133] }, "quality": "13/13 exact prompts and cached_tokens=0; historical 100-event compatibility figure 126.72926582199506 tok/s is not used as conventional decode", "evidence": "data/laguna-shared-elementwise-m12-record-20260731.json" }, { "id": "laguna-s-historical-width12-dflash11-record", "state": "lab-measured", "revision": "laguna-s-2.1-int4-4bbfc28", "variant": "INT4 target + INT4 DFlash draft with runtime E4M3FN W8A16 draft projections", "runtime": "vLLM XPU e596ef1543466ae1a05e5bb8091f58872e2b18ba + XPU kernels 6f9dd3c3a7b1b677a992ca4f431a968408f9c816", "config": {"record_order": 1, "cards": 4, "tp": 4, "ep": 4, "dflash": 11, "verifier_width": 12, "graph": "PIECEWISE", "capture_size": 12, "kv": "bfloat16"}, "workload": "sealed 13-prompt, one-start, cache-zero historical record gate with one active generation; conventional 99-interval metric", "metrics": { "decode_tok_s": [101.94172124017027], "ttft_ms": [5758.738295000512], "draft_acceptance_rate": [0.26820724334708174], "effective_tokens_per_verification": [3.950279676817899] }, "quality": "13/13 token and output-text equality and cached_tokens=0; approved 102.97143559613157 tok/s receipt used a legacy 100-event/99-interval convention and is not the plotted conventional value", "evidence": "data/laguna-s-2.1-width12-dflash-fp8-record-20260726.json" } ], "series_measurements": [ { "id": "laguna-s-tp4-dflash11-long-context-screen", "state": "lab-screened", "revision": "laguna-s-2.1-int4-4bbfc28", "variant": "INT4 target + INT4 DFlash draft / BF16 KV", "runtime": "vLLM XPU 1a7f61feffbc61b21b73f812d231c7426386ccdc + XPU kernels 99886d783372e621941228250091dc8ebdc1595d", "profile_id": "laguna-q12-long-context-gpu080-composed", "measurement_class": "composed-long-context-retrieval-screen", "promotion_status": "not-promotable-q1-mismatch-from-4k", "quality_scope": "intrinsic retrieval and cache-zero passed at every depth; target-only q1 output equality passed only at 1K", "config": { "cards": 4, "tp": 4, "ep": 4, "speculative_method": "DFlash", "draft_depth": 11, "verifier_width": 12, "graph": "PIECEWISE 146/145 target + 14/13 draft", "kv": "bfloat16", "configured_max_context_tokens": 32768, "output_tokens": 128, "max_num_batched_tokens": 8192, "gpu_memory_utilization": 0.8, "prefix_caching": false, "cache_state": "cached_tokens=0" }, "workload": "exact input-token arrays at six prompt lengths, 128 generated tokens, one active request, BF16 KV, prefix caching off; medians across early/middle/late retrieval placements except the 1K first-live capture/JIT row was excluded and two rows remained; the profile is composed from three sealed service runs", "axis": "prompt_tokens", "points": [ {"x": 1024, "decode_tok_s": 153.6043924981438, "prefill_tok_s": 5171.877711197638, "ttft_ms": 201.77322600011393, "draft_acceptance_rate": 0.23564213564213565}, {"x": 4096, "decode_tok_s": 80.24333572713317, "prefill_tok_s": 7331.793132902693, "ttft_ms": 564.3018349996964, "draft_acceptance_rate": 0.11363636363636363}, {"x": 8192, "decode_tok_s": 46.80992228390887, "prefill_tok_s": 4129.893501143041, "ttft_ms": 1993.1577850002213, "draft_acceptance_rate": 0.023402340234023402}, {"x": 16384, "decode_tok_s": 40.413761445523036, "prefill_tok_s": 5111.21405464208, "ttft_ms": 3227.641733999917, "draft_acceptance_rate": 0.008620689655172414}, {"x": 24576, "decode_tok_s": 39.782690532944315, "prefill_tok_s": 5053.232558316487, "ttft_ms": 4882.659674000024, "draft_acceptance_rate": 0.008547008547008548}, {"x": 32640, "decode_tok_s": 39.58863491052164, "prefill_tok_s": 7345.07048562943, "ttft_ms": 4477.591568999742, "draft_acceptance_rate": 0.0047430830039525695} ], "quality": "All 18 long rows returned exact prompt IDs, cached_tokens=0, 128 completion tokens, and the requested retrieval fact. Target-only q1 comparison matched all three 1K outputs but mismatched all 15 outputs from 4K through 32,640 after 67–107 common tokens; the 1K aggregate excludes the first-live capture/JIT row. This is screened evidence, not an exact or promotion-grade context curve.", "evidence": "data/laguna-s-2.1-xpu-b70/long-context-baseline-gpu080-swap24g-20260802.json" } ], "views": [ { "id": "laguna-s-record-progression", "title": "Qualified record progression", "subtitle": "Same pinned INT4 target and DFlash draft, TP4+EP4, DFlash11, width-12 verifier, BF16 KV; runtime revisions differ and values use conventional 99-interval accounting", "x_label": "qualified record generation", "discrete": true, "metrics": ["decode_tok_s", "ttft_ms"], "series": [ {"label": "qualified records", "measurement_ids": ["laguna-s-historical-width12-dflash11-record", "laguna-s-current-width12-dflash11-record"], "x_from": "config.record_order"} ] }, { "id": "laguna-s-historical-dflash-efficiency", "title": "Historical DFlash efficiency evidence", "subtitle": "The sealed 2026-07-26 packet is the only stored family measurement with aggregate acceptance and emitted tokens per draft cycle", "x_label": "draft tokens", "discrete": true, "metrics": ["draft_acceptance_rate", "effective_tokens_per_verification"], "series": [ {"label": "historical sealed packet", "measurement_ids": ["laguna-s-historical-width12-dflash11-record"], "x_from": "config.dflash"} ] } ], "coverage_views": [ { "id": "laguna-context-by-tp", "label": "long context · screened", "row_axis": { "key": "prompt_tokens", "label": "prompt", "prefix": "", "value_labels": { "1024": "1K", "4096": "4K", "8192": "8K", "16384": "16K", "24576": "24K", "32640": "32,640" } }, "column_axis": {"key": "tp", "label": "TP", "prefix": "TP"}, "fixed_selectors": { "revision": "laguna-s-2.1-int4-4bbfc28", "variant": "INT4 target + INT4 DFlash draft / BF16 KV", "runtime": "vLLM XPU 1a7f61feffbc61b21b73f812d231c7426386ccdc + XPU kernels 99886d783372e621941228250091dc8ebdc1595d", "profile_id": "laguna-q12-long-context-gpu080-composed", "ep": 4, "speculative_method": "DFlash", "draft_depth": 11, "verifier_width": 12, "kv": "bfloat16", "configured_max_context_tokens": 32768, "output_tokens": 128 }, "fixed": "Pinned INT4 target and DFlash draft · BF16 KV · q12 verifier · one request · cache zero. TP4 points are bounded retrieval-screen medians, not exact-output measurements; TP1, TP2, and TP3 were not run.", "decision_note": "Every depth passed intrinsic retrieval, but target-only q1 equality passed only at 1K and failed from 4K onward. Screened points are excluded from measured graphs and do not replace the 125.462 short-context record.", "rows": [1024, 4096, 8192, 16384, 24576, 32640], "columns": [1, 2, 3, 4], "cells": { "1024:1": {"state": "missing", "label": "not measured"}, "1024:2": {"state": "missing", "label": "not measured"}, "1024:3": {"state": "missing", "label": "not measured"}, "1024:4": {"state": "lab-screened", "label": "D153.6 · P5,172 · T201.77 · AR0.24", "evidence_id": "laguna-s-tp4-dflash11-long-context-screen", "point_x": 1024, "reason": "Two-row median after excluding the first-live capture/JIT row; retrieval and q1 output equality passed; DFlash acceptance 23.56%."}, "4096:1": {"state": "missing", "label": "not measured"}, "4096:2": {"state": "missing", "label": "not measured"}, "4096:3": {"state": "missing", "label": "not measured"}, "4096:4": {"state": "lab-screened", "label": "D80.24 · P7,332 · T564.3 · AR0.11", "evidence_id": "laguna-s-tp4-dflash11-long-context-screen", "point_x": 4096, "reason": "Three-row median; retrieval passed but q1 output equality failed; DFlash acceptance 11.36%."}, "8192:1": {"state": "missing", "label": "not measured"}, "8192:2": {"state": "missing", "label": "not measured"}, "8192:3": {"state": "missing", "label": "not measured"}, "8192:4": {"state": "lab-screened", "label": "D46.81 · P4,130 · T1,993 · AR0.02", "evidence_id": "laguna-s-tp4-dflash11-long-context-screen", "point_x": 8192, "reason": "Three-row median; retrieval passed but q1 output equality failed; DFlash acceptance 2.34%."}, "16384:1": {"state": "missing", "label": "not measured"}, "16384:2": {"state": "missing", "label": "not measured"}, "16384:3": {"state": "missing", "label": "not measured"}, "16384:4": {"state": "lab-screened", "label": "D40.41 · P5,111 · T3,228 · AR0.01", "evidence_id": "laguna-s-tp4-dflash11-long-context-screen", "point_x": 16384, "reason": "Three-row median; retrieval passed but q1 output equality failed; DFlash acceptance 0.86%."}, "24576:1": {"state": "missing", "label": "not measured"}, "24576:2": {"state": "missing", "label": "not measured"}, "24576:3": {"state": "missing", "label": "not measured"}, "24576:4": {"state": "lab-screened", "label": "D39.78 · P5,053 · T4,883 · AR0.01", "evidence_id": "laguna-s-tp4-dflash11-long-context-screen", "point_x": 24576, "reason": "Three-row median; retrieval passed but q1 output equality failed; DFlash acceptance 0.85%."}, "32640:1": {"state": "missing", "label": "not measured"}, "32640:2": {"state": "missing", "label": "not measured"}, "32640:3": {"state": "missing", "label": "not measured"}, "32640:4": {"state": "lab-screened", "label": "D39.59 · P7,345 · T4,478 · AR0", "evidence_id": "laguna-s-tp4-dflash11-long-context-screen", "point_x": 32640, "reason": "Three-row median; retrieval passed but q1 output equality failed; DFlash acceptance 0.47%."} } } ], "family_closures": [ { "selectors": {"revision": "laguna-s-2.1-int4-4bbfc28", "cards": [1, 2, 3]}, "state": "missing", "reason": "No qualified performance or quality profile is stored for this exact target/draft pair outside the four-card TP4+EP4 topology.", "evidence": "packages/laguna-s-2.1-int4-b70-125tps/package.json" }, { "selectors": {"revision": "laguna-s-2.1-int4-4bbfc28", "profile": "q1-exact-context-sweep"}, "state": "missing", "reason": "A bounded TP4 retrieval screen covers 1K through 32,640, but every 4K-and-longer output diverged from the target-only q1 oracle; no promotion-grade exact context curve is stored.", "evidence": "data/laguna-s-2.1-xpu-b70/long-context-baseline-gpu080-swap24g-20260802.json" }, { "selectors": {"packet": "laguna-s-2.1-int4-b70-125tps-20260731", "replay": "portable-runtime-build"}, "state": "missing", "reason": "The current record remains tied to exact originating-host runtime and native artifacts; a portable runtime rebuild is not stored.", "evidence": "packages/laguna-s-2.1-int4-b70-125tps/package.json" }, { "selectors": {"packet": "laguna-s-2.1-int4-b70-125tps-20260731", "replay": "non-originating-host"}, "state": "missing", "reason": "No non-originating-host replay has completed the full record gate.", "evidence": "repro/laguna-s-2.1-int4-b70-125tps-20260731/README.md" } ], "estimates": [] }