189 meaningful target-only vLLM XPU cells: three runtime-compatible artifacts × MTP0 × seven exact active-context depths × off/full+piecewise/piecewise graph × f16/E4M3/E5M2 KV. MTP depths are not blindly attached to FP8 artifacts without an exact speculator/package binding.

63/189

189 exact cells63 classified0 measured126 gaps189 retry-tagged

126 · untested63 ⚠ quarantined

Fixed: runtime_family=vLLM XPU · runtime_profile_id=qwen36-tp1-vllm-xpu-target-only-v1 · speculation_method=none

Break down by axis

Weight revisionqwen3.6-27b 63/189

Quantized artifactAutoRound INT4 · webhie 0/63FP8 · runtime-native base 0/63FP8 · historical export 63/63

Tensor parallelTP1 63/189

MTP depthMTP0 63/189

Active context0 9/272K 9/274K 9/278K 9/2716K 9/2724K 9/2732K 9/27

Graph modeoff 21/63full + piecewise 21/63piecewise 21/63

KV cacheF16 21/63FP8 E4M3 21/63FP8 E5M2 21/63

Qwen3.6 TP1 · vLLM XPU AutoRound MTP matrix

252 meaningful native-MTP cells: the measured AutoRound artifact × MTP1–4 × seven exact active-context depths × off/full+piecewise/piecewise graph × f16/E4M3/E5M2 KV. Other artifacts stay outside this profile until their speculator binding is established.

0/252

252 exact cells0 classified0 measured252 gaps252 retry-tagged

252 · untested

Fixed: runtime_family=vLLM XPU · runtime_profile_id=qwen36-tp1-vllm-xpu-autoround-native-mtp-v1 · speculation_method=native-mtp

Break down by axis

Weight revisionqwen3.6-27b 0/252

Quantized artifactAutoRound INT4 · webhie 0/252

Tensor parallelTP1 0/252

MTP depthMTP1 0/63MTP2 0/63MTP3 0/63MTP4 0/63

Active context0 0/362K 0/364K 0/368K 0/3616K 0/3624K 0/3632K 0/36

Graph modeoff 0/84full + piecewise 0/84piecewise 0/84

KV cacheF16 0/84FP8 E4M3 0/84FP8 E5M2 0/84

Qwen3.6 TP1 · llama.cpp SYCL MTP matrix

448 meaningful llama.cpp SYCL speculative cells: four MTP-bearing GGUF artifacts × MTP1–4 × seven exact active-context depths × graph off/SYCL × f16/q8_0 KV. MTP0 is represented in the separate target-only profile; vLLM compilation modes and FP8 KV are excluded.

56/448

448 exact cells56 classified28 measured392 gaps448 retry-tagged

28 ✓ measured28 ◇ screened392 · untested

Fixed: runtime_family=llama.cpp SYCL · runtime_profile_id=qwen36-tp1-llamacpp-sycl-embedded-mtp-v1 · speculation_method=embedded-mtp

Break down by axis

Weight revisionqwen3.6-27b 56/448

Quantized artifactQ4_0 · unsloth MTP 0/112UD-Q4_K_XL · unsloth MTP 0/112Q4_K_M · unsloth MTP 0/112Q8_0 · unsloth embedded MTP 56/112

Tensor parallelTP1 56/448

MTP depthMTP1 14/112MTP2 14/112MTP3 14/112MTP4 14/112

Active context0 8/642K 8/644K 8/648K 8/6416K 8/6424K 8/6432K 8/64

Graph modeoff 56/224SYCL graph 0/224

KV cacheF16 28/224Q8_0 28/224

Qwen3.6 1 card · llama.cpp SYCL, no draft model

168 meaningful target-only cells: all six Qwen3.6 GGUF target artifacts × MTP0 × seven exact active-context depths × graph off/SYCL × f16/q8_0 KV. Intrinsic MTP1–4 is never manufactured for target-only files; those depths live only in the separately bound MTP-bearing profile.

112/168

168 exact cells112 classified112 measured56 gaps168 retry-tagged

112 ✓ measured56 · untested

Fixed: runtime_family=llama.cpp SYCL · runtime_profile_id=qwen36-tp1-llamacpp-sycl-target-only-v1 · speculation_method=none

Break down by axis

Weight revisionqwen3.6-27b 112/168

Quantized artifactQ4_0 · unsloth MTP 14/28UD-Q4_K_XL · unsloth MTP 14/28Q4_K_M · unsloth MTP 28/28Q8_0 · unsloth embedded MTP 28/28Q8_0 · unsloth target 28/28Q8_0 · ggml-org 0/28

Tensor parallelTP1 112/168

MTP depthMTP0 112/168

Active context0 16/242K 16/244K 16/248K 16/2416K 16/2424K 16/2432K 16/24

Graph modeoff 70/84SYCL graph 42/84

KV cacheF16 56/84Q8_0 56/84

Qwen3.8 1 card · vLLM XPU, no draft model

126 meaningful target-only vLLM XPU cells: two runtime-compatible artifacts × MTP0 × seven exact active-context depths × off/full+piecewise/piecewise graph × f16/E4M3/E5M2 KV. MTP depths are not blindly attached to the official FP8 artifact without an exact speculator/package binding.

44/126

126 exact cells44 classified32 measured82 gaps126 retry-tagged

32 ✓ measured82 · untested2 ⚠ quarantined10 × unsupported

Fixed: runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp1-vllm-xpu-target-only-v1 · speculation_method=none

Break down by axis

Weight revisionqwen3.8-27b 44/126

Quantized artifactAutoRound INT4 · devan-carlin 38/63FP8 · official 6/63

Tensor parallelTP1 44/126

MTP depthMTP0 44/126

Active context0 1/182K 7/184K 8/188K 7/1816K 7/1824K 7/1832K 7/18

Graph modeoff 25/42full + piecewise 7/42piecewise 12/42

KV cacheF16 24/42FP8 E4M3 13/42FP8 E5M2 7/42

Qwen3.8 TP1 · vLLM XPU AutoRound MTP matrix

252 meaningful native-MTP cells: the measured AutoRound artifact × MTP1–4 × seven exact active-context depths × off/full+piecewise/piecewise graph × f16/E4M3/E5M2 KV. The official FP8 artifact stays outside this profile until its exact speculator binding is established.

21/252

252 exact cells21 classified15 measured231 gaps252 retry-tagged

15 ✓ measured231 · untested6 ⚠ quarantined

Fixed: runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp1-vllm-xpu-autoround-native-mtp-v1 · speculation_method=native-mtp

Break down by axis

Weight revisionqwen3.8-27b 21/252

Quantized artifactAutoRound INT4 · devan-carlin 21/252

Tensor parallelTP1 21/252

MTP depthMTP1 8/63MTP2 7/63MTP3 6/63MTP4 0/63

Active context0 0/362K 3/364K 6/368K 3/3616K 3/3624K 3/3632K 3/36

Graph modeoff 19/84full + piecewise 0/84piecewise 2/84

KV cacheF16 20/84FP8 E4M3 1/84FP8 E5M2 0/84

Qwen3.8 TP1 · current f01e AutoRound eager MTP4 screened depth

Seven exact current-f01e TP1 native-MTP4 eager F16-KV selectors: 4K/16K/24K are Grade D lab-screened only, 2K and conflicting separate-boot 8K are quarantined, 32K is runtime-closed, and x0 is missing. There are zero lab-measured cells and structural cells expose no speed.

6/7

7 exact cells6 classified0 measured1 gaps7 retry-tagged

1 ■ closed3 ◇ screened1 · untested2 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=1 · mtp=4 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp1-vllm-xpu-autoround-native-mtp4-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 pinned e9d1398 · AutoRound decision closure

The original 96-cell TP/MTP/graph/KV campaign, pinned only to vLLM XPU nightly e9d1398d9. Measured target and TP1 eager-MTP anchors remain measured; unsupported, closed, and quarantined descendants are explicit. Nothing in this contract transfers to f01e or another runtime.

96/96

96 exact cells96 classified9 measured0 gaps96 retry-tagged

6 ■ closed9 ✓ measured33 ⚠ quarantined48 × unsupported

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · runtime=vLLM XPU nightly e9d1398d9 · configured_max_context_tokens=32768

Break down by axis

Tensor parallelTP1 24/24TP2 24/24TP3 24/24TP4 24/24

MTP depthMTP0 24/24MTP1 24/24MTP2 24/24MTP3 24/24

Graphoff 48/48on 48/48

KV cacheF16 32/32FP8 E4M3 32/32FP8 E5M2 32/32

Qwen3.8 1 card · llama.cpp SYCL, no draft model

112 meaningful target-only cells: four GGUF artifacts × MTP0 × seven exact active-context depths × graph off/SYCL × f16/q8_0 KV. These cells preserve 84 exact measurements and one speedless exact-selector closure. All 14 superseded live estimates are retired; their frozen snapshots remain data-only calibration history, and no MTP speed is transferred.

85/112

112 exact cells85 classified84 measured27 gaps112 retry-tagged

1 ■ closed84 ✓ measured27 · untested

Fixed: runtime_family=llama.cpp SYCL

Break down by axis

Weight revisionqwen3.8-27b 85/112

Quantized artifactQ4_K_M · ggml-org 21/28Q8_0 · ggml-org 22/28UD-Q5_K_S · unsloth 21/28UD-Q4_K_XL · unsloth 21/28

Tensor parallelTP1 85/112

MTP depthMTP0 85/112

Active context0 12/162K 12/164K 12/168K 13/1616K 12/1624K 12/1632K 12/16

Graph modeoff 56/56SYCL graph 29/56

KV cacheF16 56/56Q8_0 29/56

Qwen3.8 TP1 · llama.cpp SYCL MTP-package matrix

224 meaningful package cells: the two targets pinned with the inspected Unsloth Q4_0 MTP draft × MTP1–4 × seven exact active-context depths × graph off/SYCL × f16/q8_0 KV. This is an exact package backlog; no Qwen3.6 speed or quality transfers.

4/224

224 exact cells4 classified0 measured220 gaps224 retry-tagged

4 ◇ screened220 · untested

Fixed: runtime_family=llama.cpp SYCL · runtime_profile_id=qwen38-tp1-llamacpp-sycl-external-mtp-v1 · speculation_method=external-mtp-draft · speculator_artifact_id=qwen38-27b-mtp-q4-0-50d9ce5

Break down by axis

Weight revisionqwen3.8-27b 4/224

Quantized artifactUD-Q5_K_S · unsloth 4/112UD-Q4_K_XL · unsloth 0/112

Tensor parallelTP1 4/224

MTP depthMTP1 1/56MTP2 1/56MTP3 1/56MTP4 1/56

Active context0 0/322K 0/324K 0/328K 4/3216K 0/3224K 0/3232K 0/32

Graph modeoff 4/112SYCL graph 0/112

KV cacheF16 0/112Q8_0 4/112

Qwen3.8 TP2 · llama.cpp SYCL Q4_K_M exact HTTP depth

Seven exact selector cells for the pinned Q4_K_M TP2 target-only F16-KV graph-off HTTP lane: six Grade C cache-zero measurements at 2K-32K and an explicit missing x=0 cell.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-ggmlorg-q4-k-m-0669b98 · tp=2 · mtp=0 · graph_mode=off · kv=f16 · runtime_family=llama.cpp SYCL · runtime_profile_id=qwen38-tp2-llamacpp-sycl-q4km-http-depth-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · llama.cpp SYCL Q8_0 exact HTTP depth

Seven exact selector cells for the pinned Q8_0-weight TP2 target-only F16-KV graph-off HTTP lane: six Grade C cache-zero measurements at 2K-32K and an explicit missing x=0 cell.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-ggmlorg-q8-0-0669b98 · tp=2 · mtp=0 · graph_mode=off · kv=f16 · runtime_family=llama.cpp SYCL · runtime_profile_id=qwen38-tp2-llamacpp-sycl-q8-http-depth-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · vLLM XPU official FP8 exact HTTP depth

Seven exact selector cells for the pinned official-FP8 TP2 target-only F16-KV PIECEWISE HTTP lane: six Grade C cache-zero measurements at 2K-32K and an explicit missing x=0 cell.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-official-fp8-017b9c7 · tp=2 · mtp=0 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-official-fp8-http-depth-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · vLLM XPU official FP8/W8A16 deterministic MTP1 exact HTTP depth

Seven exact selector cells for the qualified official-FP8/W8A16 TP2 deterministic-MTP1 F16-KV graph-off HTTP lane: six Grade C cache-zero, target-array-exact measurements at 2K-32K and an explicit missing x=0 cell.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-official-fp8-017b9c7 · tp=2 · mtp=1 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-official-fp8-w8a16-mtp1-http-depth-r33 · speculation_method=publisher MTP depth 1

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · vLLM XPU official FP8 exact HTTP depth

Seven exact selector cells for the pinned official-FP8 TP4 target-only F16-KV PIECEWISE HTTP lane: six Grade C cache-zero measurements at 2K-32K and an explicit missing x=0 cell. This slower profile does not replace the AutoRound TP4 frontier.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-official-fp8-017b9c7 · tp=4 · mtp=0 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-official-fp8-http-depth-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · vLLM XPU AutoRound exact HTTP depth

Seven exact selector cells for the pinned AutoRound INT4 TP2 target-only MTP0 F16-KV FULL_AND_PIECEWISE HTTP lane: six Grade C cache-zero measurements at 2K-32K and an explicit missing x=0 cell.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=0 · graph_mode=FULL_AND_PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-http-depth-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · vLLM XPU AutoRound exact HTTP depth

Seven exact selector cells for the pinned AutoRound INT4 TP4 target-only MTP0 F16-KV FULL_AND_PIECEWISE HTTP lane: six Grade C cache-zero measurements at 2K-32K and an explicit missing x=0 cell.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=0 · graph_mode=FULL_AND_PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-http-depth-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · current f01e AutoRound eager target depth

Seven exact selector cells for current-f01e AutoRound INT4 TP2 target-only MTP0 eager F16-KV: six Grade C cache-zero nonzero measurements with exact TP1 comparison parity and explicit missing x=0.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=0 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-eager-target-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · current f01e AutoRound PIECEWISE partial target depth

Seven exact current-f01e AutoRound INT4 TP2 target-only MTP0 PIECEWISE F16-KV selectors: four Grade C target-identical measurements at 2K/4K/24K/32K, exact 8K/16K speedless target-divergence quarantines, and explicit missing x0. The dated FULL_AND_PIECEWISE graph contract remains separate.

6/7

7 exact cells6 classified4 measured1 gaps7 retry-tagged

4 ✓ measured1 · untested2 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=0 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-f01e-piecewise-target-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · current f01e AutoRound PIECEWISE MTP1 depth

Seven exact selector cells for current-f01e AutoRound INT4 TP2 native-MTP1 PIECEWISE F16-KV: one Grade C cache-zero exact-4K measurement with positive isolated acceptance, dual-parent token parity, full quality, graph/topology/cache/model verification, and six explicit missing depths.

1/7

7 exact cells1 classified1 measured6 gaps7 retry-tagged

1 ✓ measured6 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=1 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-native-mtp1-piecewise-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 0/14K 1/18K 0/116K 0/124K 0/132K 0/1

Qwen3.8 TP2 · current f01e AutoRound PIECEWISE MTP2 depth

Seven exact selector cells for current-f01e AutoRound INT4 TP2 native-MTP2 PIECEWISE F16-KV: one Grade C cache-zero exact-4K measurement with positive isolated acceptance, dual-parent token parity, full quality, graph/topology/cache/model verification, and six explicit missing depths.

1/7

7 exact cells1 classified1 measured6 gaps7 retry-tagged

1 ✓ measured6 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=2 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-native-mtp2-piecewise-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 0/14K 1/18K 0/116K 0/124K 0/132K 0/1

Qwen3.8 TP2 · current f01e AutoRound eager MTP1 depth

Seven exact selector cells for current-f01e AutoRound INT4 TP2 native-MTP1 eager F16-KV: six Grade C cache-zero nonzero measurements with positive isolated acceptance, exact same-topology TP2/MTP0 target parity, and explicit missing x=0.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=1 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-native-mtp1-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · current f01e AutoRound eager MTP2 partial depth

Seven exact selector cells for current-f01e AutoRound INT4 TP2 native-MTP2 eager F16-KV: five Grade C cache-zero target-parity measurements at 4K-32K, one exact/cache-zero/positive-acceptance 2K quarantine after token-90 target divergence, and explicit missing x=0.

6/7

7 exact cells6 classified5 measured1 gaps7 retry-tagged

5 ✓ measured1 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=2 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-native-mtp2-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · current f01e AutoRound eager MTP3 parent-scoped depth

Seven exact selector cells for current-f01e AutoRound INT4 TP2 native-MTP3 eager F16-KV: five Grade C cache-zero target-parity measurements at 4K-32K, one 2K cell structurally excluded by the parent MTP2 token-90 quarantine with no MTP3 speed, and explicit missing x=0.

6/7

7 exact cells6 classified5 measured1 gaps7 retry-tagged

5 ✓ measured1 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=3 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-native-mtp3-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP2 · current f01e AutoRound eager MTP4 sentinel depth

Seven exact current-f01e AutoRound INT4 TP2 native-MTP4 eager F16-KV selectors: exact 4K, 16K, and 24K are Grade C target-parity measurements, exact 8K is a structural target-parity quarantine with no site speed, and x0/2K/32K remain missing. No value is inferred from another MTP depth or topology.

4/7

7 exact cells4 classified3 measured3 gaps7 retry-tagged

3 ✓ measured3 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=2 · mtp=4 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp2-vllm-xpu-autoround-native-mtp4-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 0/14K 1/18K 1/116K 1/124K 1/132K 0/1

Qwen3.8 TP4 · current f01e AutoRound eager oracle depth

Seven exact selector cells for the current-f01e AutoRound INT4 TP4 target-only MTP0 eager F16-KV lane: six Grade C cache-zero nonzero measurements and explicit missing x=0. The earlier exact-8K cell is retained; five cells are additive from the expansion.

6/7

7 exact cells6 classified6 measured1 gaps7 retry-tagged

6 ✓ measured1 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=0 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-eager-oracle-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · current f01e AutoRound PIECEWISE partial target depth

Seven exact current-f01e AutoRound INT4 TP4 target-only MTP0 PIECEWISE F16-KV selectors: five Grade C target-identical measurements at 2K/4K/16K/24K/32K, exact 8K speedless target-divergence quarantine, and explicit missing x0. The dated FULL_AND_PIECEWISE graph contract remains separate.

6/7

7 exact cells6 classified5 measured1 gaps7 retry-tagged

5 ✓ measured1 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=0 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-f01e-piecewise-target-v1 · speculation_method=none

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · current f01e AutoRound PIECEWISE MTP1 depth

Seven exact current-f01e AutoRound INT4 TP4 native-MTP1 PIECEWISE F16-KV selectors: exact 4K is a Grade C dual-parent-parity measurement and x0/2K/8K/16K/24K/32K remain missing. The known current-profile 8K graph-corruption signature is retained as a caveat, not transferred into an unrun MTP1 cell.

1/7

7 exact cells1 classified1 measured6 gaps7 retry-tagged

1 ✓ measured6 · untested

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=1 · graph_mode=PIECEWISE · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-native-mtp1-piecewise-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 0/14K 1/18K 0/116K 0/124K 0/132K 0/1

Qwen3.8 TP4 · current f01e AutoRound eager MTP1 depth

Seven exact selector cells for current-f01e AutoRound INT4 TP4 native-MTP1 eager F16-KV: five Grade C target-parity measurements at 4K-32K, 2K quarantined at token 90, and x0 missing. The prior 8K cell is retained unchanged.

6/7

7 exact cells6 classified5 measured1 gaps7 retry-tagged

5 ✓ measured1 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=1 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-native-mtp1-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · current f01e AutoRound eager MTP2 depth

Seven exact current-f01e TP4 native-MTP2 eager F16-KV selectors: five Grade C target-parity measurements at 4K-32K, 2K quarantined at token 90, and x0 missing; prior 8K retained.

6/7

7 exact cells6 classified5 measured1 gaps7 retry-tagged

5 ✓ measured1 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=2 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-native-mtp2-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · current f01e AutoRound eager MTP3 depth

Seven exact current-f01e TP4 native-MTP3 eager F16-KV selectors: five Grade C target-parity measurements at 4K-32K, 2K quarantined at token 90, and x0 missing; prior 8K retained.

6/7

7 exact cells6 classified5 measured1 gaps7 retry-tagged

5 ✓ measured1 · untested1 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=3 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-native-mtp3-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · current f01e AutoRound eager MTP4 depth

Seven exact current-f01e TP4 native-MTP4 eager F16-KV selectors: 4K/16K/24K are Grade C target-parity measurements, 2K/8K are quarantined target divergences, 32K is a runtime-fatal closure, and x0 is missing. Structural cells expose no speed.

6/7

7 exact cells6 classified3 measured1 gaps7 retry-tagged

1 ■ closed3 ✓ measured1 · untested2 ⚠ quarantined

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · tp=4 · mtp=4 · graph_mode=off · kv=f16 · runtime_family=vLLM XPU · runtime_profile_id=qwen38-tp4-vllm-xpu-autoround-native-mtp4-v1 · speculation_method=native-mtp

Break down by axis

Active context0 0/12K 1/14K 1/18K 1/116K 1/124K 1/132K 1/1

Qwen3.8 TP4 · vLLM XPU AutoRound fully checked snapshot

One exact short-suite TP4 anchor for the frozen b2dd9ce73d + 1e90ffa672 source stack. This is not an active-context curve: 32K is configured capacity only, and no x=0/2K/4K/8K/16K/24K/32K throughput cell is claimed.

1/1

1 exact cells1 classified1 measured0 gaps1 retry-tagged

1 ✓ measured

Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · quantization=AutoRound INT4 · runtime_family=vLLM XPU · runtime=vLLM XPU b2dd9ce73d + XPU kernels 1e90ffa672 · measurement_class=strict-natural-eos · workload=two strict natural-EOS replays; 25 unique cold prompts; up to 512 output tokens; cached_tokens=0; conventional 99-interval median · tp=4 · mtp=0 · graph=on · graph_mode=FULL_AND_PIECEWISE · kv=f16 · configured_max_context_tokens=32768 · gpu_memory_utilization=0.6 · natural_eos=True

Break down by axis

Frozen source stackb2dd9ce73d + 1e90ffa672 1/1

Qwen3.6 Q4_0 MTP × TP

Qwen3.6 Q4_0 intrinsic-MTP artifact · llama.cpp SYCL · graph off · q8_0 target KV · fixed cold/cache-zero suite. Every MTP0–5 × TP1/2/3/4 cell is explicit. Fixed: revision=qwen3.6-27b · artifact_id=qwen36-27b-unsloth-mtp-q4-0-20c9c45 · quantization=Q4_0 · runtime_family=llama.cpp SYCL · profile_id=q36-q4-0-tp1-intrinsic-mtp-v1 · graph=off · kv=q8_0.

✓ Measured1 card, no speculative decoding TP1·MTP025.68 tok/sreport · evidence
✓ Measured1 card, + speculative decoding (depth 3) TP1·MTP348.43 tok/sreport · evidence

22 untested combinations

TP2·MTP0TP3·MTP0TP4·MTP0TP1·MTP1TP2·MTP1TP3·MTP1TP4·MTP1TP1·MTP2TP2·MTP2TP3·MTP2TP4·MTP2TP2·MTP3TP3·MTP3TP4·MTP3TP1·MTP4TP2·MTP4TP3·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP3·MTP5TP4·MTP5

Full matrix and exact selectors

Qwen3.6 UD-Q4_K_XL MTP × TP

Qwen3.6 UD-Q4_K_XL · llama.cpp SYCL · intrinsic MTP with n_min=0/p_min=0 · f16 KV · graph requested but replay unproven · fixed cold/cache-zero suite. Fixed: revision=qwen3.6-27b · artifact_id=qwen36-27b-unsloth-mtp-ud-q4-k-xl-4085665 · quantization=UD-Q4_K_XL · runtime_family=llama.cpp SYCL · profile_id=q36-udq4kxl-tp1-intrinsic-mtp-depth-v1 · graph=requested-not-proven-replay · kv=f16.

✓ Measured1 card, no speculative decoding TP1·MTP023.44 tok/sreport · evidence
✓ Measured1 card, + speculative decoding (depth 3) TP1·MTP329.22 tok/sreport · evidence
✓ Measured1 card, + speculative decoding (depth 4) TP1·MTP428.31 tok/sreport · evidence
✓ Measured1 card, + speculative decoding (depth 5) TP1·MTP524.66 tok/sreport · evidence

20 untested combinations

TP2·MTP0TP3·MTP0TP4·MTP0TP1·MTP1TP2·MTP1TP3·MTP1TP4·MTP1TP1·MTP2TP2·MTP2TP3·MTP2TP4·MTP2TP2·MTP3TP3·MTP3TP4·MTP3TP2·MTP4TP3·MTP4TP4·MTP4TP2·MTP5TP3·MTP5TP4·MTP5

Full matrix and exact selectors

Qwen3.6 Q4_0 speculation × TP

Qwen3.6 Q4_0 target on llama.cpp SYCL with graph off and q8_0 target KV. Rows distinguish target-only, bundled intrinsic MTP3, and a separate native Q8_0 DFlash5 draft. Fixed: revision=qwen3.6-27b · artifact_id=qwen36-27b-unsloth-mtp-q4-0-20c9c45 · quantization=Q4_0 · runtime_family=llama.cpp SYCL · graph=off · kv=q8_0.

✓ Measured1 card, target only TP1·target only25.68 tok/sreport · evidence
✓ Measured1 card, intrinsic MTP3 TP1·intrinsic MTP348.43 tok/sreport · evidence
✓ Measured1 card, native DFlash5 TP1·native DFlash547.34 tok/sreport · evidence

9 untested combinations

TP2·target onlyTP3·target onlyTP4·target onlyTP2·intrinsic MTP3TP3·intrinsic MTP3TP4·intrinsic MTP3TP2·native DFlash5TP3·native DFlash5TP4·native DFlash5

Full matrix and exact selectors

Qwen3.6 runtime FP8 × TP

Qwen3.6 base weights quantized at runtime through the vLLM native-FP8 path, replayed in the reference lab with 262,144 configured max, eager vLLM XPU, and fp8_e4m3 KV. This is not a prequantized FP8 checkpoint. Fixed: revision=qwen3.6-27b · artifact_id=qwen36-27b-base-runtime-fp8-6a9e13b · runtime_family=vLLM XPU · mtp=0 · graph=off · kv=fp8_e4m3 · configured_max_context_tokens=262144.

✓ Measured2 cards, FP8 TP2·FP830.17 tok/sreport · evidence

3 untested combinations

TP1·FP8TP3·FP8TP4·FP8

Full matrix and exact selectors

Qwen3.6 Q8 MTP × TP

Qwen3.6 Q8_0 llama.cpp SYCL evidence with graph disabled and f16 KV. Every MTP0–5 × TP1/2/3/4 cell is explicit; independent replicas are not mislabeled as tensor parallel. Fixed: revision=qwen3.6-27b · quantization=Q8_0 · runtime_family=llama.cpp SYCL · graph=off · kv=f16.

✓ Measured1 card, no speculative decoding TP1·MTP017.11 tok/sevidence
✓ Measured2 cards, no speculative decoding TP2·MTP036.6 tok/sreport · evidence
✓ Measured1 card, + speculative decoding (depth 3) TP1·MTP336.05 tok/sreport · evidence

21 untested combinations

TP3·MTP0TP4·MTP0TP1·MTP1TP2·MTP1TP3·MTP1TP4·MTP1TP1·MTP2TP2·MTP2TP3·MTP2TP4·MTP2TP2·MTP3TP3·MTP3TP4·MTP3TP1·MTP4TP2·MTP4TP3·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP3·MTP5TP4·MTP5

Full matrix and exact selectors

Qwen3.6 Q8 speculation × serving slots

Qwen3.6 Q8_0 TP1 llama.cpp SYCL at 32K configured capacity per slot. The target c1 cell uses official isolated two-prompt band evidence; target-only c2 was actually launched near 32K, while embedded MTP3 c2 stopped at the safe-fit gate. Fixed: revision=qwen3.6-27b · quantization=Q8_0 · runtime_family=llama.cpp SYCL · tp=1 · graph=off · kv=f16 · context_tokens_per_slot=32768.

✓ Measuredc1, target only c1·target onlyD15.077 @17K / 13.686 @31.8K · D511evidence
✓ Measuredc2, target only c2·target only10.14 tok/sevidence
✓ Measuredc1, embedded MTP3 c1·embedded MTP3D44.49 @17K / 41.49 @31.8Kevidence
■ Closedc2, embedded MTP3 c2·embedded MTP3 — A second target and draft KV plus recurrent state exceed safe B70 capacity before useful headroom; retry only with a materially lower-memory, separately preregistered identity.32,683 MiB fit no-goevidence

Codes in the rows: D = decode tok/s · P = prefill tok/s · T = ms to first token · AR = share of drafted tokens accepted · A = combined tok/s.

Full matrix and exact selectors

Qwen3.6 AutoRound MTP × TP

Qwen3.6 Webhie AutoRound INT4 with the BF16-scale INT8 target LM-head, graph enabled, and f16 KV. Cells show normalized historical evidence already in the lab; they do not imply that dated runtimes and suites are directly rankable. Fixed: revision=qwen3.6-27b · artifact_id=qwen36-27b-autoround-w4a16-f5750c9 · quantization=AutoRound INT4 · runtime_family=vLLM XPU · graph=on · kv=f16.

✓ Measured1 card, + speculative decoding (depth 1) TP1·MTP151.25 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 2) TP1·MTP259.59 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 3) TP1·MTP368.24 tok/sevidence
✓ Measured2 cards, + speculative decoding (depth 3) TP2·MTP395.38 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 4) TP1·MTP459.89 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 5) TP1·MTP5 — Both strict/cache-zero MTP5 capture-size arms are recorded; neither beat the MTP3 control, so the depth/config-only lane is closed.59.82 tok/sevidence
× Unsupported3 cards, no speculative decoding TP3·MTP0 — Qwen3.6 and Qwen3.8 share this architecture; the vLLM tensor-parallel route cannot evenly partition 16 GDN key heads across TP3.16 heads cannot split by 3evidence
× Unsupported3 cards, + speculative decoding (depth 1) TP3·MTP1 — Qwen3.6 and Qwen3.8 share this architecture; the vLLM tensor-parallel route cannot evenly partition 16 GDN key heads across TP3.16 heads cannot split by 3evidence
× Unsupported3 cards, + speculative decoding (depth 2) TP3·MTP2 — Qwen3.6 and Qwen3.8 share this architecture; the vLLM tensor-parallel route cannot evenly partition 16 GDN key heads across TP3.16 heads cannot split by 3evidence
× Unsupported3 cards, + speculative decoding (depth 3) TP3·MTP3 — Qwen3.6 and Qwen3.8 share this architecture; the vLLM tensor-parallel route cannot evenly partition 16 GDN key heads across TP3.16 heads cannot split by 3evidence
× Unsupported3 cards, + speculative decoding (depth 4) TP3·MTP4 — Qwen3.6 and Qwen3.8 share this architecture; the vLLM tensor-parallel route cannot evenly partition 16 GDN key heads across TP3.16 heads cannot split by 3evidence
× Unsupported3 cards, + speculative decoding (depth 5) TP3·MTP5 — Qwen3.6 and Qwen3.8 share this architecture; the vLLM tensor-parallel route cannot evenly partition 16 GDN key heads across TP3.16 heads cannot split by 3evidence

12 untested combinations

TP1·MTP0TP2·MTP0TP4·MTP0TP2·MTP1TP4·MTP1TP2·MTP2TP4·MTP2TP4·MTP3TP2·MTP4TP4·MTP4TP2·MTP5TP4·MTP5

Full matrix and exact selectors

pinned e9d1398 · graph off

Qwen3.8 AutoRound · pinned vLLM XPU e9d1398 · f16 KV · short fixed-output diagnostic. TP2/TP4 MTP1/3 were closed by the frozen MTP2 expansion gate; MTP4/5 were outside the 96-cell program and remain missing.

✓ Measured1 card, no speculative decoding TP1·MTP024.25 tok/sevidence
✓ Measured2 cards, no speculative decoding TP2·MTP016.77 tok/sevidence
✓ Measured4 cards, no speculative decoding TP4·MTP017.38 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 1) TP1·MTP14.51 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 2) TP1·MTP24.41 tok/sevidence
✓ Measured4 cards, + speculative decoding (depth 2) TP4·MTP231.17 tok/sevidence
✓ Measured1 card, + speculative decoding (depth 3) TP1·MTP34.3 tok/sevidence
◇ Speed check only2 cards, + speculative decoding (depth 2) TP2·MTP2boot+canary onlyevidence
■ Closed2 cards, + speculative decoding (depth 1) TP2·MTP1gateevidence
■ Closed4 cards, + speculative decoding (depth 1) TP4·MTP1gateevidence
■ Closed2 cards, + speculative decoding (depth 3) TP2·MTP3MTP2 gateevidence
■ Closed4 cards, + speculative decoding (depth 3) TP4·MTP3MTP2 gateevidence
× Unsupported3 cards, no speculative decoding TP3·MTP0head splitevidence
× Unsupported3 cards, + speculative decoding (depth 1) TP3·MTP1head splitevidence
× Unsupported3 cards, + speculative decoding (depth 2) TP3·MTP2head splitevidence
× Unsupported3 cards, + speculative decoding (depth 3) TP3·MTP3head splitevidence
× Unsupported3 cards, + speculative decoding (depth 4) TP3·MTP4head splitevidence
× Unsupported3 cards, + speculative decoding (depth 5) TP3·MTP5head splitevidence

6 untested combinations

TP1·MTP4TP2·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP4·MTP5

Full matrix and exact selectors

pinned e9d1398 · graph on

Qwen3.8 AutoRound · pinned vLLM XPU e9d1398 · f16 KV · short fixed-output diagnostic. Multi-GPU XPU Graph is runtime-labeled unsupported/experimental; the e9d1398 TP1 MTP1 corruption anchor quarantined MTP1/2/3 in the bounded program. MTP4/5 remain missing.

✓ Measured1 card, no speculative decoding TP1·MTP030.26 tok/sevidence
✓ Measured2 cards, no speculative decoding TP2·MTP048.95 tok/sevidence
✓ Measured4 cards, no speculative decoding TP4·MTP071.67 tok/sevidence
× Unsupported3 cards, no speculative decoding TP3·MTP0head splitevidence
× Unsupported3 cards, + speculative decoding (depth 1) TP3·MTP1head splitevidence
× Unsupported3 cards, + speculative decoding (depth 2) TP3·MTP2head splitevidence
× Unsupported3 cards, + speculative decoding (depth 3) TP3·MTP3head splitevidence
× Unsupported3 cards, + speculative decoding (depth 4) TP3·MTP4head splitevidence
× Unsupported3 cards, + speculative decoding (depth 5) TP3·MTP5head splitevidence
⚠ Quarantined1 card, + speculative decoding (depth 1) TP1·MTP1 — output not usable. Observed: 7.63 tok/s.evidence
⚠ Quarantined2 cards, + speculative decoding (depth 1) TP2·MTP1 — output not usable. Observed: graph+MTP.evidence
⚠ Quarantined4 cards, + speculative decoding (depth 1) TP4·MTP1 — output not usable. Observed: graph+MTP.evidence
⚠ Quarantined1 card, + speculative decoding (depth 2) TP1·MTP2 — output not usable. Observed: MTP1 anchor.evidence
⚠ Quarantined2 cards, + speculative decoding (depth 2) TP2·MTP2 — output not usable. Observed: graph+MTP.evidence
⚠ Quarantined4 cards, + speculative decoding (depth 2) TP4·MTP2 — output not usable. Observed: graph+MTP.evidence
⚠ Quarantined1 card, + speculative decoding (depth 3) TP1·MTP3 — output not usable. Observed: MTP1 anchor.evidence
⚠ Quarantined2 cards, + speculative decoding (depth 3) TP2·MTP3 — output not usable. Observed: graph+MTP.evidence
⚠ Quarantined4 cards, + speculative decoding (depth 3) TP4·MTP3 — output not usable. Observed: graph+MTP.evidence

6 untested combinations

TP1·MTP4TP2·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP4·MTP5

Full matrix and exact selectors

dated current-main snapshots × TP

Qwen3.8 AutoRound · exact dated vLLM/XPU-kernel source stacks · strict natural-EOS profile · MTP0 · f16 KV · graph on. Each row is immutable historical evidence, never a moving 'current' alias. Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · variant=AutoRound INT4 · runtime_family=vLLM XPU · measurement_class=strict-natural-eos · mtp=0 · graph=on · kv=f16 · configured_max_context_tokens=32768 · natural_eos=True.

✓ Measured1 card, 0ecc + stock kernel TP1·0ecc + stock kernel30.33 tok/sevidence
✓ Measured1 card, 0ecc + baaa TP1·0ecc + baaa30.28 tok/sevidence
✓ Measured1 card, 6a9 + baaa TP1·6a9 + baaa30.27 tok/sevidence
✓ Measured1 card, b2dd + 1e90 TP1·b2dd + 1e9030.28 tok/sevidence
✓ Measured4 cards, b2dd + 1e90 TP4·b2dd + 1e90 — All model and quality arms passed; deterministic aggregation was recovered offline after a mawk reserved-name wrapper failure.71.83 tok/sevidence
× Unsupported3 cards, 0ecc + stock kernel TP3·0ecc + stock kernelhead split
× Unsupported3 cards, 0ecc + baaa TP3·0ecc + baaahead split
× Unsupported3 cards, 6a9 + baaa TP3·6a9 + baaahead split
× Unsupported3 cards, b2dd + 1e90 TP3·b2dd + 1e90head split

7 untested combinations

TP2·0ecc + stock kernelTP4·0ecc + stock kernelTP2·0ecc + baaaTP4·0ecc + baaaTP2·6a9 + baaaTP4·6a9 + baaaTP2·b2dd + 1e90

Full matrix and exact selectors

stock a3561ef8 · graph on

Qwen3.8 AutoRound · stock vLLM XPU a3561ef8 · strict natural-EOS profile · f16 KV · graph on. No MTP1–5 value is inferred from e9d1398; diagnostic captures remain in their separate graph. Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · variant=AutoRound INT4 · runtime=vLLM XPU nightly a3561ef8 stock · profile_id=a3561ef8-stock · measurement_class=strict-natural-eos · graph=on · kv=f16 · configured_max_context_tokens=32768 · natural_eos=True.

✓ Measured1 card, no speculative decoding TP1·MTP030.24 tok/sevidence
✓ Measured2 cards, no speculative decoding TP2·MTP048.49 tok/sevidence
✓ Measured4 cards, no speculative decoding TP4·MTP0 — The 71.900 strict capture passed quality but did not replicate and is not a promoted record.71.9 tok/sevidence
× Unsupported3 cards, no speculative decoding TP3·MTP0head splitevidence
× Unsupported3 cards, + speculative decoding (depth 1) TP3·MTP1head splitevidence
× Unsupported3 cards, + speculative decoding (depth 2) TP3·MTP2head splitevidence
× Unsupported3 cards, + speculative decoding (depth 3) TP3·MTP3head splitevidence
× Unsupported3 cards, + speculative decoding (depth 4) TP3·MTP4head splitevidence
× Unsupported3 cards, + speculative decoding (depth 5) TP3·MTP5head splitevidence

15 untested combinations

TP1·MTP1TP2·MTP1TP4·MTP1TP1·MTP2TP2·MTP2TP4·MTP2TP1·MTP3TP2·MTP3TP4·MTP3TP1·MTP4TP2·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP4·MTP5

Full matrix and exact selectors

a3561ef8 TP2 overlay · graph on

Qwen3.8 AutoRound · vLLM XPU a3561ef8 plus the TP2 historical-winner decisions · strict natural-EOS profile · MTP0 · f16 KV · graph on. This is measured partial recovery and not promoted. Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · variant=AutoRound INT4 · runtime=vLLM XPU nightly a3561ef8 + TP2 historical-winner overlay · profile_id=a3561ef8-tp2-e9d1398-best-config-overlay-v1 · measurement_class=strict-natural-eos · graph=on · kv=f16 · configured_max_context_tokens=32768 · natural_eos=True.

✓ Measured2 cards, no speculative decoding TP2·MTP0 — Full quality passed, but strict speed missed the frozen 49.019651 gate by 0.021%; not promoted.49.01 tok/sevidence
× Unsupported3 cards, no speculative decoding TP3·MTP0head split
× Unsupported3 cards, + speculative decoding (depth 1) TP3·MTP1head split
× Unsupported3 cards, + speculative decoding (depth 2) TP3·MTP2head split
× Unsupported3 cards, + speculative decoding (depth 3) TP3·MTP3head split
× Unsupported3 cards, + speculative decoding (depth 4) TP3·MTP4head split
× Unsupported3 cards, + speculative decoding (depth 5) TP3·MTP5head split

17 untested combinations

TP1·MTP0TP4·MTP0TP1·MTP1TP2·MTP1TP4·MTP1TP1·MTP2TP2·MTP2TP4·MTP2TP1·MTP3TP2·MTP3TP4·MTP3TP1·MTP4TP2·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP4·MTP5

Full matrix and exact selectors

a3561ef8 TP4 overlay · graph on

Qwen3.8 AutoRound · vLLM XPU a3561ef8 plus the TP4 historical-winner decisions · strict natural-EOS profile · MTP0 · f16 KV · graph on. This is the accepted versioned overlay. Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-autoround-w4a16-bce40ca · variant=AutoRound INT4 · runtime=vLLM XPU nightly a3561ef8 + TP4 historical-winner overlay · profile_id=a3561ef8-tp4-e9d1398-best-config-overlay-v1 · measurement_class=strict-natural-eos · graph=on · kv=f16 · configured_max_context_tokens=32768 · natural_eos=True.

✓ Measured4 cards, no speculative decoding TP4·MTP0 — Accepted observed strict range; neither endpoint is independently replicated.71.45 tok/sevidence
× Unsupported3 cards, no speculative decoding TP3·MTP0head split
× Unsupported3 cards, + speculative decoding (depth 1) TP3·MTP1head split
× Unsupported3 cards, + speculative decoding (depth 2) TP3·MTP2head split
× Unsupported3 cards, + speculative decoding (depth 3) TP3·MTP3head split
× Unsupported3 cards, + speculative decoding (depth 4) TP3·MTP4head split
× Unsupported3 cards, + speculative decoding (depth 5) TP3·MTP5head split

17 untested combinations

TP1·MTP0TP2·MTP0TP1·MTP1TP2·MTP1TP4·MTP1TP1·MTP2TP2·MTP2TP4·MTP2TP1·MTP3TP2·MTP3TP4·MTP3TP1·MTP4TP2·MTP4TP4·MTP4TP1·MTP5TP2·MTP5TP4·MTP5

Full matrix and exact selectors

Q4_K_M context × TP

Qwen3.8 Q4_K_M · llama.cpp SYCL · target only · f16 KV. Exact raw-engine decode/prefill sweep points are shown from the measured zero-depth anchor through 32K. Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-ggmlorg-q4-k-m-0669b98 · variant=Q4_K_M · runtime=llama.cpp SYCL · mtp=0 · graph=off · kv=f16.

✓ Measured1 card, 0 TP1·0D24.81 · P825.24guide · evidence
✓ Measured1 card, 2K TP1·2KD24.46 · P919.67guide · evidence
✓ Measured1 card, 4K TP1·4KD24.25 · P892.64guide · evidence
✓ Measured1 card, 8K TP1·8KD23.83 · P850.97guide · evidence
✓ Measured1 card, 16K TP1·16KD23.1 · P779.5guide · evidence
✓ Measured1 card, 24K TP1·24KD22.42 · P719.39guide · evidence
✓ Measured1 card, 32K TP1·32KD21.77 · P667.83guide · evidence

Codes in the rows: D = decode tok/s · P = prefill tok/s · T = ms to first token · AR = share of drafted tokens accepted · A = combined tok/s.

21 untested combinations

TP2·0TP3·0TP4·0TP2·2KTP3·2KTP4·2KTP2·4KTP3·4KTP4·4KTP2·8KTP3·8KTP4·8KTP2·16KTP3·16KTP4·16KTP2·24KTP3·24KTP4·24KTP2·32KTP3·32KTP4·32K

Full matrix and exact selectors

Q4_K_M q8 KV context × TP

Qwen3.8 Q4_K_M · llama.cpp SYCL · target only · q8_0 KV. Exact raw-engine decode/prefill sweep points are shown from the measured zero-depth anchor through 32K. Fixed: revision=qwen3.8-27b · artifact_id=qwen38-27b-ggmlorg-q4-k-m-0669b98 · variant=Q4_K_M · runtime=llama.cpp SYCL · mtp=0 · graph=off · kv=q8_0.

✓ Measured1 card, 0 TP1·0D24.27 · P817.78guide · evidence
✓ Measured1 card, 2K TP1·2KD22.45 · P912.17guide · evidence
✓ Measured1 card, 4K TP1·4KD21.05 · P887.17guide · evidence
✓ Measured1 card, 8K TP1·8KD18.68 · P843.21guide · evidence
✓ Measured1 card, 16K TP1·16KD14.86 · P772.01guide · evidence
✓ Measured1 card, 24K TP1·24KD12.4 · P711.48guide · evidence
✓ Measured1 card, 32K TP1·32KD10.66 · P662.56guide · evidence

Codes in the rows: D = decode tok/s · P = prefill tok/s · T = ms to first token · AR = share of drafted tokens accepted · A = combined tok/s.

21 untested combinations

TP2·0TP3·0TP4·0TP2·2KTP3·2KTP4·2KTP2·4KTP3·4KTP4·4KTP2·8KTP3·8KTP4·8KTP2·16KTP3·16KTP4·16KTP2·24KTP3·24KTP4·24KTP2·32KTP3·32KTP4·32K

Full matrix and exact selectors

256K package quant × TP

Qwen3.8 full-package TP1 serving profile with q8_0 target/draft KV, MTP3, and mmproj-F16 resident. 262144 is configured capacity; the measured serving prompts were 48–78 tokens, not 256K active context. Fixed: revision=qwen3.8-27b · runtime_family=llama.cpp SYCL · profile_id=q38-256k-vision-mtp3-q8kv-v1 · measurement_class=conventional-99-interval-cold-realistic-suite · mtp=3 · graph=off · kv=q8_0 · configured_max_context_tokens=262144 · vision=mmproj-F16 resident.

✓ Measured1 card, UD-Q5_K_S TP1·UD-Q5_K_S26.67 tok/sguide · evidence
✓ Measured1 card, UD-Q4_K_XL TP1·UD-Q4_K_XL27.51 tok/sevidence

6 untested combinations

TP2·UD-Q5_K_STP3·UD-Q5_K_STP4·UD-Q5_K_STP2·UD-Q4_K_XLTP3·UD-Q4_K_XLTP4·UD-Q4_K_XL

Full matrix and exact selectors

llama.cpp quant × TP

Qwen3.8 target-only llama.cpp/SYCL routes with f16 KV. Cells retain their recorded workload identity: some TP1 quant rows expose raw context curves, while Q4_K_M TP2 and Q8_0 TP1/TP2 expose conventional or strict varied-prompt suite headlines. Fixed: revision=qwen3.8-27b · runtime=llama.cpp SYCL · mtp=0 · graph=off · kv=f16.

✓ Measured1 card, Q4_K_M TP1·Q4_K_Mmeasured context curveguide · evidence
✓ Measured2 cards, Q4_K_M TP2·Q4_K_M49.72 tok/sguide · evidence
✓ Measured1 card, UD-Q5_K_S TP1·UD-Q5_K_Smeasured context curveevidence
✓ Measured1 card, UD-Q4_K_XL TP1·UD-Q4_K_XLmeasured context curveevidence
✓ Measured1 card, Q8_0 TP1·Q8_019.62 tok/sguide · evidence
✓ Measured2 cards, Q8_0 TP2·Q8_036.73 tok/sguide · evidence

10 untested combinations

TP3·Q4_K_MTP4·Q4_K_MTP2·UD-Q5_K_STP3·UD-Q5_K_STP4·UD-Q5_K_STP2·UD-Q4_K_XLTP3·UD-Q4_K_XLTP4·UD-Q4_K_XLTP3·Q8_0TP4·Q8_0

Full matrix and exact selectors

vLLM quant × TP

Qwen3.8 target-only vLLM XPU routes with f16 KV and graph enabled. AutoRound cells use the stock a3561ef8 strict profile; FP8 uses its separately pinned official-model runtime. Fixed: revision=qwen3.8-27b · runtime_family=vLLM XPU · mtp=0 · graph=on · kv=f16.

✓ Measured1 card, AutoRound INT4 TP1·AutoRound INT430.24 tok/sreport · evidence
✓ Measured2 cards, AutoRound INT4 TP2·AutoRound INT448.49 tok/sreport · evidence
✓ Measured4 cards, AutoRound INT4 TP4·AutoRound INT4 — The 71.900 capture did not replicate and is not promoted.71.9 tok/sreport · evidence
✓ Measured2 cards, FP8 TP2·FP821.71 tok/sguide · evidence
× Unsupported3 cards, AutoRound INT4 TP3·AutoRound INT4head split
× Unsupported3 cards, FP8 TP3·FP8head split

2 untested combinations

TP1·FP8TP4·FP8

Full matrix and exact selectors

Measured results

Every number links to its proof.

family data

What changes with card count

Qwen3.8 AutoRound · pinned vLLM XPU e9d1398 · MTP0 · f16 KV · fixed-output diagnostic; TP2/TP4 XPU Graph is runtime-labeled unsupported/experimental. No a3561ef8 eager, prefill, or TTFT values are inferred here.

pinned e9d1398 eagerpinned e9d1398 graphunsupported

pinned e9d1398 eager 1: 23.72–24.25 · 2: 16.77 · 4: 17.38
pinned e9d1398 graph 1: 30.22–30.26 · 2: 48.83–48.95 · 4: 71.55–71.67

All measured metric summaries

Decode

pinned e9d1398 eager 1: 23.72–24.25 · 2: 16.77 · 4: 17.38
pinned e9d1398 graph 1: 30.22–30.26 · 2: 48.83–48.95 · 4: 71.55–71.67

Prefill

pinned e9d1398 eager 1: 281.5–281.7 · 2: 495.3 · 4: 865.5
pinned e9d1398 graph 1: 276.3–277.9 · 2: 509 · 4: 849.5–864.8

TTFT

pinned e9d1398 eager 1: 274.5–274.6 · 2: 162 · 4: 89.8
pinned e9d1398 graph 1: 277.5–278.6 · 2: 150.31–152.2 · 4: 90.8–91.1

evidence 1 · evidence 2

Draft depth (one pinned build)

Qwen3.8 AutoRound · TP1 eager · f16 KV · pinned vLLM XPU e9d1398; preregistered ladder ended at MTP3. No a3561ef8 MTP value is inferred.

pinned e9d1398 eager

pinned e9d1398 eager 0: 23.72–24.25 · 1: 4.51 · 2: 4.41 · 3: 4.3

All measured metric summaries

Decode

pinned e9d1398 eager 0: 23.72–24.25 · 1: 4.51 · 2: 4.41 · 3: 4.3

Acceptance length

pinned e9d1398 eager 1: 1.91 · 2: 2.7 · 3: 3.47

Prefill

pinned e9d1398 eager 0: 281.5–281.7 · 1: 168.2 · 2: 199.4 · 3: 161.3

TTFT

pinned e9d1398 eager 0: 274.5–274.6 · 1: 487.5 · 2: 411.8 · 3: 518.7

evidence 1

Q4_K_M HTTP context × KV/graph

Qwen3.8 Q4_K_M · TP1 llama.cpp · graph-off F16 is Grade C cache-zero HTTP at 2K-32K and keeps x=0 as a separate raw point · exact cache64 graph-patched F16 covers 0-32K · Q8_0 remains the preserved raw-engine graph-off curve · profiles remain separately labeled

F16 KV · graph off · HTTP 2K-32K · Grade CF16 KV · SYCL graph cache64 · HTTP 0-32K · Grade CQ8_0 KV · graph off · raw 0-32K

F16 KV · graph off · HTTP 2K-32K · Grade C 2K: 27.26 · 4K: 27.19 · 8K: 26.87 · 16K: 26 · 24K: 25.24 · 32K: 24.49
F16 KV · SYCL graph cache64 · HTTP 0-32K · Grade C 0: 26.72 · 2K: 26.27 · 4K: 26.02 · 8K: 25.54 · 16K: 24.58 · 24K: 23.74 · 32K: 23.22
Q8_0 KV · graph off · raw 0-32K 0: 24.27 · 2K: 22.45 · 4K: 21.05 · 8K: 18.68 · 16K: 14.86 · 24K: 12.4 · 32K: 10.66

evidence 1 · evidence 2 · evidence 3

Conversation length × compression

Qwen3.8 · TP1 llama.cpp · f16 KV · separately pinned builds and raw-engine shape sweeps; inspect each series identity before cross-quant comparison

Q4_K_MUD-Q5_K_SUD-Q4_K_XLQ8_0 weights

Q4_K_M 0: 24.81 · 2K: 24.46 · 4K: 24.25 · 8K: 23.83 · 16K: 23.1 · 24K: 22.42 · 32K: 21.77
UD-Q5_K_S 0: 22.72 · 2K: 22.37 · 4K: 22.13 · 8K: 21.79 · 16K: 21.19 · 24K: 20.61 · 32K: 20.09
UD-Q4_K_XL 0: 21.81 · 2K: 21.53 · 4K: 21.37 · 8K: 21.06 · 16K: 20.49 · 24K: 19.95 · 32K: 19.45
Q8_0 weights 0: 19.66 · 2K: 19.59 · 4K: 19.51 · 8K: 19.29 · 16K: 18.84 · 24K: 18.42 · 32K: 18.02

All measured metric summaries

Decode

Prefill

Q4_K_M 0: 825.24 · 2K: 919.67 · 4K: 892.64 · 8K: 850.97 · 16K: 779.5 · 24K: 719.39 · 32K: 667.83
UD-Q5_K_S 0: 849.3 · 2K: 778.35 · 4K: 759.39 · 8K: 726.59 · 16K: 671.59 · 24K: 625.83 · 32K: 588.21
UD-Q4_K_XL 0: 762.57 · 2K: 735.45 · 4K: 719.47 · 8K: 689.58 · 16K: 639.46 · 24K: 596.49 · 32K: 564.86
Q8_0 weights 0: 996.89 · 2K: 987.39 · 4K: 959.56 · 8K: 914.06 · 16K: 837.51 · 24K: 773.15 · 32K: 719.14

evidence 1 · evidence 2 · evidence 3

33 more evidence views

Card count (fully checked runs)

Qwen3.8 AutoRound · MTP0 · f16 KV · graph on · strict natural-EOS. Pinned e9d1398, stock a3561ef8, and a3561ef8 winner overlays remain separate; stock TP4 71.900 is a captured high, not a replicated promotion; TP2 overlay is partial and TP4 overlay is accepted.

pinned e9d1398 strictstock a3561ef8 stricta3561ef8 overlay · TP2 partial / TP4 acceptedunsupported

pinned e9d1398 strict 1: 30.31 · 2: 49.02 · 4: 71.29–71.4
stock a3561ef8 strict 1: 30.24–30.24 · 2: 48.49 · 4: 71.25–71.9
a3561ef8 overlay · TP2 partial / TP4 accepted 2: 49.01 · 4: 71.35–71.45

evidence 1 · evidence 2 · evidence 3 · evidence 4

Card count (quick checks)

Qwen3.8 AutoRound · MTP0 · f16 KV · graph on · fixed-output ignore-EOS diagnostic. These are measured diagnostic captures, not strict deployment headlines; the overlay was measured only at TP2/TP4.

pinned e9d1398 diagnosticstock a3561ef8 diagnostica3561ef8 overlay diagnosticunsupported

pinned e9d1398 diagnostic 1: 30.22–30.26 · 2: 48.83–48.95 · 4: 71.55–71.67
stock a3561ef8 diagnostic 1: 30.33 · 2: 48.65 · 4: 71.34
a3561ef8 overlay diagnostic 2: 49.06 · 4: 71.72

evidence 1 · evidence 2 · evidence 3 · evidence 4 · evidence 5

Qwen3.6 draft-depth ladder

Webhie AutoRound · TP1 · graph capture 8 · same-window strict gate · historical generated-tokens-1–100 accounting; MTP0 was not rerun in this window

Qwen3.6 same-window

Qwen3.6 same-window 1: 51.25 · 2: 59.59 · 3: 64.73 · 4: 59.89

All measured metric summaries

Decode

Qwen3.6 same-window 1: 51.25 · 2: 59.59 · 3: 64.73 · 4: 59.89

TTFT

Qwen3.6 same-window 1: 298.76 · 2: 437.35 · 3: 610.37 · 4: 779.48

evidence 1

Qwen3.6 draft depth 5 (quick check)

Webhie AutoRound · TP1 · MTP5 · PIECEWISE graph · same-window strict/cache-zero screen · historical generated-tokens-1–100 accounting; neither capture size beat MTP3.

MTP5 · max graph capture size 859.26tok/sDecode985.84msTTFThistorical 100 event compatibility · strict valid depth screen no win · evidence

MTP5 · max graph capture size 1659.82tok/sDecode986.62msTTFThistorical 100 event compatibility · strict valid capture screen no win · evidence

MTP5

evidence 1

Qwen3.6 draft depth 3: then vs. now

AutoRound · TP1/TP2 · graph on · f16 KV. Historical highs and dated current support use disclosed different suites/accounting; whiskers are captured support bands, not a regression claim.

historical 12-prompt highsdated current supportunsupported

historical 12-prompt highs 1: 68.24 · 2: 95.38
dated current support 1: 65.36–66.72 · 2: 94.32–94.79

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.6 Q4_0 with its built-in draft

TP1 llama.cpp SYCL · exact Q4_0 artifact · graph off · q8_0 KV · fixed 12-prompt cold/cache-zero suite · MTP3 is 48.434 conventional (48.923285 retained helper); MTP1/2/4/5 remain gaps

target only → intrinsic MTP3 · draft tokens 025.68tok/sDecode1,147msTTFTconventional 99 interval recomputed from historical helper · matched target only control · evidence

target only → intrinsic MTP3 · draft tokens 348.43tok/sDecode0.59ratioDraft acceptance2.76tokensAcceptance length1,179msTTFTconventional 99 interval recomputed from historical helper · qualified intrinsic mtp3 result · evidence

target only → intrinsic MTP3missing

evidence 1 · evidence 2

Qwen3.6 Q4_0 with a DFlash draft

TP1 native DFlash5 · Q4_0 target + Q8_0 draft · configuration-exact fusion-off/on A/B, not output-exact · confirmed candidate is +8.14% versus control at 47.341 conventional (47.818818 retained helper)

same-config DFlash5 A/B · Q6_K draft-head top-1 fusion (0 off · 1 on) 043.78tok/sDecode1,154msTTFTconventional 99 interval recomputed from historical helper · same config dflash5 fusion off control · evidence

same-config DFlash5 A/B · Q6_K draft-head top-1 fusion (0 off · 1 on) 146.64–47.34tok/sDecode1,157msTTFTconventional 99 interval recomputed from historical helper · confirmed dflash5 fusion record · evidence

same-config DFlash5 A/B

evidence 1 · evidence 2

Qwen3.6 UD-Q4_K_XL draft depth

Four independent TP1 services in one matched wave · MTP0/3/4/5 · n_min=0 · p_min=0 · fixed 12-prompt cold/cache-zero suite · graph requested but replay unproven · conventional 99-interval accounting

matched p_min=0 depth wavemissing

matched p_min=0 depth wave 0: 23.44 · 3: 29.22 · 4: 28.31 · 5: 24.66

All measured metric summaries

Decode

matched p_min=0 depth wave 0: 23.44 · 3: 29.22 · 4: 28.31 · 5: 24.66

Draft acceptance

matched p_min=0 depth wave 3: 0.59 · 4: 0.51 · 5: 0.43

TTFT

matched p_min=0 depth wave 0: 411.04 · 3: 423.16 · 4: 422.72 · 5: 423.87

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.6 UD-Q4_K_XL deep draft (closed)

TP1 policy screen · each point changes n_min and/or p_min as labeled · discrete closure, not a pure depth curve · best is 31.165 conventional (31.480049 retained helper) · fixed cold/cache-zero suite

MTP5 · n1 · p.75MTP7 · n1 · p.75MTP7 · n1 · p.65MTP9 · n2 · p.75missing

MTP5 · n1 · p.75 5: 30.73
MTP7 · n1 · p.75 7: 30.68
MTP7 · n1 · p.65 7: 31.17
MTP9 · n2 · p.75 9: 27.04

All measured metric summaries

Decode

MTP5 · n1 · p.75 5: 30.73
MTP7 · n1 · p.75 7: 30.68
MTP7 · n1 · p.65 7: 31.17
MTP9 · n2 · p.75 9: 27.04

TTFT

MTP5 · n1 · p.75 5: 421
MTP7 · n1 · p.75 7: 427.97
MTP7 · n1 · p.65 7: 424.65
MTP9 · n2 · p.75 9: 427.32

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.8 by software build

Separate older TP2 runtime family · honest MTP5 three-arm median 101.17 (100.46–101.39) · only 21–22/25 self-agreement; MTP1–4 omitted, not interpolated

older source stack · draft tokens 048.76tok/sDecodelab measurement; quality scope in evidence · evidence

older source stack · draft tokens 5100.45–101.39tok/sDecodelab measurement; quality scope in evidence · evidence

older source stack

evidence 1

Qwen3.6 quant × KV context

TP1 llama.cpp SYCL · MTP0 · graph off · raw-engine exact-depth points only; quantizations and KV types remain separate series

Q4_K_M · f16 KVQ4_K_M · q8_0 KVUD-Q4_K_XL · f16 KVUD-Q4_K_XL · q8_0 KVQ4_0 · f16 KVQ4_0 · q8_0 KVQ8_0 embedded-MTP artifact · f16 KVQ8_0 embedded-MTP artifact · q8_0 KVQ8_0 target-only Unsloth · f16 KVQ8_0 target-only Unsloth · q8_0 KV

Q4_K_M · f16 KV 0: 29.3 · 2K: 28.72 · 4K: 28.42 · 8K: 27.88 · 16K: 26.88 · 24K: 25.9 · 32K: 25.04
Q4_K_M · q8_0 KV 0: 28.43 · 2K: 25.91 · 4K: 24 · 8K: 20.92 · 16K: 16.17 · 24K: 13.3 · 32K: 11.38
UD-Q4_K_XL · f16 KV 0: 28.2 · 2K: 27.73 · 4K: 27.4 · 8K: 26.9 · 16K: 25.97 · 24K: 25.11 · 32K: 24.31
UD-Q4_K_XL · q8_0 KV 0: 27.4 · 2K: 25.05 · 4K: 23.21 · 8K: 20.33 · 16K: 15.86 · 24K: 13.11 · 32K: 11.21
Q4_0 · f16 KV 0: 26.4 · 2K: 25.61 · 4K: 24.98 · 8K: 23.89 · 16K: 21.96 · 24K: 20.31 · 32K: 18.91
Q4_0 · q8_0 KV 0: 25.87 · 2K: 23.94 · 4K: 22.51 · 8K: 20.21 · 16K: 16.84 · 24K: 14.42 · 32K: 12.6
Q8_0 embedded-MTP artifact · f16 KV 0: 19.83 · 2K: 19.64 · 4K: 19.5 · 8K: 19.27 · 16K: 18.81 · 24K: 18.39 · 32K: 17.98
Q8_0 embedded-MTP artifact · q8_0 KV 0: 19.41 · 2K: 18.27 · 4K: 17.34 · 8K: 15.71 · 16K: 12.88 · 24K: 10.96 · 32K: 9.57
Q8_0 target-only Unsloth · f16 KV 0: 19.84 · 2K: 19.64 · 4K: 19.51 · 8K: 19.27 · 16K: 18.81 · 24K: 18.38 · 32K: 17.98
Q8_0 target-only Unsloth · q8_0 KV 0: 19.41 · 2K: 18.27 · 4K: 17.35 · 8K: 15.71 · 16K: 12.88 · 24K: 10.96 · 32K: 9.57

All measured metric summaries

Decode

Prefill

Q4_K_M · f16 KV 0: 826.36 · 2K: 891.6 · 4K: 869.41 · 8K: 830.46 · 16K: 760.02 · 24K: 704.98 · 32K: 655.1
Q4_K_M · q8_0 KV 0: 824.62 · 2K: 884.56 · 4K: 865.76 · 8K: 823.1 · 16K: 756.11 · 24K: 698.56 · 32K: 648.26
UD-Q4_K_XL · f16 KV 0: 840.76 · 2K: 890.41 · 4K: 864.38 · 8K: 826.18 · 16K: 758.86 · 24K: 701.5 · 32K: 654.62
UD-Q4_K_XL · q8_0 KV 0: 837.72 · 2K: 883.05 · 4K: 859.58 · 8K: 822.39 · 16K: 751.83 · 24K: 696.25 · 32K: 645.98
Q4_0 · f16 KV 0: 635.64 · 2K: 300.21 · 4K: 284.81 · 8K: 255.79 · 16K: 203.91 · 24K: 168.95 · 32K: 144.91
Q4_0 · q8_0 KV 0: 634.03 · 2K: 299.3 · 4K: 284.23 · 8K: 256.07 · 16K: 203.93 · 24K: 168.64 · 32K: 144.96
Q8_0 embedded-MTP artifact · f16 KV 0: 908.1 · 2K: 897.5 · 4K: 871.16 · 8K: 830.35 · 16K: 763.17 · 24K: 707.36 · 32K: 658.01
Q8_0 embedded-MTP artifact · q8_0 KV 0: 901.98 · 2K: 890.55 · 4K: 863.3 · 8K: 823.43 · 16K: 753.25 · 24K: 701.09 · 32K: 653.03
Q8_0 target-only Unsloth · f16 KV 0: 907.79 · 2K: 897.08 · 4K: 872.57 · 8K: 830.75 · 16K: 762.67 · 24K: 708.32 · 32K: 660.85
Q8_0 target-only Unsloth · q8_0 KV 0: 901.91 · 2K: 891.93 · 4K: 862.37 · 8K: 823.05 · 16K: 757.05 · 24K: 702.84 · 32K: 653.73

evidence 1 · evidence 2 · evidence 3 · evidence 4 · evidence 5 · evidence 6 · evidence 7 · evidence 8 · evidence 9 · evidence 10

Q4_K_XL HTTP context × KV/graph

Grade C · same Qwen3.8 UD-Q4_K_XL artifact · target-only TP1 · exact cache-zero HTTP serving through 32K · graph-off F16/Q8_0 and exact cache64 graph-patched F16 profiles remain separately labeled · all full quality batteries passed

F16 KV · graph off · Grade CF16 KV · SYCL graph cache64 · Grade CQ8_0 KV · graph off · Grade C

F16 KV · graph off · Grade C 0: 21.83 · 2K: 21.31 · 4K: 20.87 · 8K: 20.02 · 16K: 18.64 · 24K: 17.37 · 32K: 16.39
F16 KV · SYCL graph cache64 · Grade C 0: 23.21 · 2K: 22.82 · 4K: 22.66 · 8K: 22.27 · 16K: 21.45 · 24K: 20.76 · 32K: 20.35
Q8_0 KV · graph off · Grade C 0: 21.77 · 2K: 20.27 · 4K: 19.09 · 8K: 17.12 · 16K: 13.83 · 24K: 11.65 · 32K: 10.08

evidence 1 · evidence 2 · evidence 3

Q8_0-weight HTTP context × KV/graph

Grade C · same Qwen3.8 Q8_0 artifact · target-only TP1 · graph-off F16/Q8_0 and cache64 graph-patched F16 remain separately labeled · the graph profile can be slower and does not replace graph-off · all full quality batteries passed

F16 KV · graph off · HTTP · Grade CF16 KV · SYCL graph cache64 · HTTP · Grade CQ8_0 KV · graph off · HTTP · Grade C

F16 KV · graph off · HTTP · Grade C 0: 15.72 · 2K: 15.46 · 4K: 15.23 · 8K: 14.78 · 16K: 14.01 · 24K: 13.32 · 32K: 12.68
F16 KV · SYCL graph cache64 · HTTP · Grade C 0: 19.17 · 2K: 18.98 · 4K: 18.93 · 8K: 18.69 · 16K: 18.22 · 24K: 17.77 · 32K: 17.52
Q8_0 KV · graph off · HTTP · Grade C 0: 15.64 · 2K: 14.87 · 4K: 14.24 · 8K: 13.14 · 16K: 11.1 · 24K: 9.64 · 32K: 8.55

evidence 1 · evidence 2 · evidence 3

Q5_K_S HTTP context × KV/graph

Grade C · same Qwen3.8 UD-Q5_K_S artifact · target-only TP1 · exact cache-zero HTTP serving through 32K · graph-off F16/Q8_0 and exact cache64 graph-patched F16 profiles remain separately labeled · all full quality batteries passed

F16 KV · graph off · Grade CF16 KV · SYCL graph cache64 · Grade CQ8_0 KV · graph off · Grade C

F16 KV · graph off · Grade C 0: 22.62 · 2K: 22.07 · 4K: 21.51 · 8K: 20.7 · 16K: 19.12 · 24K: 17.85 · 32K: 16.73
F16 KV · SYCL graph cache64 · Grade C 0: 23.99 · 2K: 23.64 · 4K: 23.43 · 8K: 23 · 16K: 22.22 · 24K: 21.51 · 32K: 21.02
Q8_0 KV · graph off · Grade C 0: 22.49 · 2K: 20.88 · 4K: 19.59 · 8K: 17.5 · 16K: 14.05 · 24K: 11.82 · 32K: 10.21

evidence 1 · evidence 2 · evidence 3

Qwen3.8 TP2 HTTP context × format

Grade C · exact cache-zero HTTP serving at 2K-32K for three separately pinned artifacts and runtimes; x=0 remains missing · Q4_K_M and Q8_0 use llama.cpp graph-off while official FP8 uses vLLM PIECEWISE · cross-format comparisons are descriptive, not causal, and quality scopes remain distinct

Q4_K_M · F16 KV · llama.cpp graph offQ8_0 weights · F16 KV · llama.cpp graph offOfficial FP8 · F16 KV · vLLM PIECEWISE

Q4_K_M · F16 KV · llama.cpp graph off 2K: 49.49 · 4K: 49.01 · 8K: 48.3 · 16K: 47.03 · 24K: 45.53 · 32K: 44.44
Q8_0 weights · F16 KV · llama.cpp graph off 2K: 36.8 · 4K: 36.55 · 8K: 36.01 · 16K: 35.09 · 24K: 34.52 · 32K: 33.85
Official FP8 · F16 KV · vLLM PIECEWISE 2K: 21.84 · 4K: 21.67 · 8K: 21.27 · 16K: 20.93 · 24K: 20.65 · 32K: 20.39

All measured metric summaries

Decode

Prefill

Q8_0 weights · F16 KV · llama.cpp graph off 2K: 1,032 · 4K: 1,039 · 8K: 1,020 · 16K: 983.52 · 24K: 947.33 · 32K: 915.09

TTFT

Q4_K_M · F16 KV · llama.cpp graph off 2K: 1,945 · 4K: 3,860 · 8K: 7,862 · 16K: 16,300 · 24K: 25,347 · 32K: 35,059
Q8_0 weights · F16 KV · llama.cpp graph off 2K: 2,003 · 4K: 3,957 · 8K: 8,047 · 16K: 16,681 · 24K: 25,958 · 32K: 35,832
Official FP8 · F16 KV · vLLM PIECEWISE 2K: 1,385 · 4K: 2,606 · 8K: 5,192 · 16K: 10,533 · 24K: 16,139 · 32K: 21,873

evidence 1 · evidence 2 · evidence 3

Official FP8 TP1/TP2/TP4 HTTP context

Grade C · same exact official Qwen3.8 FP8 weights · TP1 eager fit supports measured 2K-8K; target-only TP2/TP4 PIECEWISE and the separately qualified TP2 W8A16/MTP1 graph-off profile reach measured 2K-32K · x=0 and unmeasured TP1 depths remain missing · profiles are separately labeled and not a causal scaling A/B

Official FP8 · TP1 · F16 KV · eagerOfficial FP8 · TP2 · F16 KV · vLLM PIECEWISEOfficial FP8 + W8A16 · TP2 · MTP1 · deterministic graph offOfficial FP8 · TP4 · F16 KV · vLLM PIECEWISE

Official FP8 · TP1 · F16 KV · eager 2K: 10.75 · 4K: 10.76 · 8K: 10.74
Official FP8 · TP2 · F16 KV · vLLM PIECEWISE 2K: 21.84 · 4K: 21.67 · 8K: 21.27 · 16K: 20.93 · 24K: 20.65 · 32K: 20.39
Official FP8 + W8A16 · TP2 · MTP1 · deterministic graph off 2K: 44.78 · 4K: 54.93 · 8K: 51.31 · 16K: 51.29 · 24K: 43.72 · 32K: 46.64
Official FP8 · TP4 · F16 KV · vLLM PIECEWISE 2K: 35.53 · 4K: 34.9 · 8K: 33.97 · 16K: 33.64 · 24K: 33.27 · 32K: 33.11

All measured metric summaries

Decode

TTFT

Official FP8 · TP1 · F16 KV · eager 2K: 2,289 · 4K: 4,592 · 8K: 9,333
Official FP8 · TP2 · F16 KV · vLLM PIECEWISE 2K: 1,385 · 4K: 2,606 · 8K: 5,192 · 16K: 10,533 · 24K: 16,139 · 32K: 21,873
Official FP8 + W8A16 · TP2 · MTP1 · deterministic graph off 2K: 799.1 · 4K: 1,187 · 8K: 2,343 · 16K: 4,869 · 24K: 7,586 · 32K: 10,487
Official FP8 · TP4 · F16 KV · vLLM PIECEWISE 2K: 1,973 · 4K: 3,663 · 8K: 7,200 · 16K: 14,424 · 24K: 21,792 · 32K: 29,236

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.8 AutoRound TP1 current-image KV/graph

Grade C · exact cache-zero HTTP serving · same current AutoRound weights and official f01e/ac7509e2 image · TP1 · MTP0 · F16 PIECEWISE plots target-identical 2K/4K/16K/24K/32K with 8K quarantined · F16 and FP8 E4M3 KV remain separately labeled · x=0 remains missing

F16 KV · eagerFP8 E4M3 KV · eagerF16 KV · PIECEWISE parity-qualified partialFP8 E4M3 KV · PIECEWISE

F16 KV · eager 2K: 11.92 · 4K: 12.12 · 8K: 11.92 · 16K: 12.11 · 24K: 12.04 · 32K: 12.11
FP8 E4M3 KV · eager 2K: 12.11 · 4K: 11.99 · 8K: 12.09 · 16K: 12.18 · 24K: 12.16 · 32K: 12.16
F16 KV · PIECEWISE parity-qualified partial 2K: 30.08 · 4K: 29.41 · 16K: 28.19 · 24K: 27.46 · 32K: 26.76
FP8 E4M3 KV · PIECEWISE 2K: 29.76 · 4K: 28.94 · 8K: 28.66 · 16K: 28.03 · 24K: 27.32 · 32K: 26.78

All measured metric summaries

Decode

TTFT

F16 KV · eager 2K: 1,520 · 4K: 2,758 · 8K: 5,645 · 16K: 11,681 · 24K: 18,097 · 32K: 24,928
FP8 E4M3 KV · eager 2K: 1,532 · 4K: 2,805 · 8K: 5,817 · 16K: 12,452 · 24K: 19,854 · 32K: 27,973
F16 KV · PIECEWISE parity-qualified partial 2K: 1,363 · 4K: 2,475 · 16K: 10,545 · 24K: 16,544 · 32K: 22,672
FP8 E4M3 KV · PIECEWISE 2K: 1,383 · 4K: 2,535 · 8K: 5,269 · 16K: 11,340 · 24K: 18,166 · 32K: 25,756

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.8 AutoRound TP1 graph modes

Grade C · exact cache-zero HTTP serving · same AutoRound weights · MTP0 · F16 KV · current f01e eager spans 2K-32K, current f01e PIECEWISE plots target-identical 2K/4K/16K/24K/32K with 8K quarantined, and FULL_AND_PIECEWISE remains the separately labeled b2dd/1e90 profile · x=0 remains missing

eager · f01e/ac7509e2PIECEWISE parity-qualified partial · f01e/ac7509e2FULL_AND_PIECEWISE · b2dd/1e90

eager · f01e/ac7509e2 2K: 11.92 · 4K: 12.12 · 8K: 11.92 · 16K: 12.11 · 24K: 12.04 · 32K: 12.11
PIECEWISE parity-qualified partial · f01e/ac7509e2 2K: 30.08 · 4K: 29.41 · 16K: 28.19 · 24K: 27.46 · 32K: 26.76
FULL_AND_PIECEWISE · b2dd/1e90 2K: 30.1 · 4K: 29.77 · 8K: 29.28 · 16K: 28.42 · 24K: 27.66 · 32K: 26.99

All measured metric summaries

Decode

TTFT

eager · f01e/ac7509e2 2K: 1,520 · 4K: 2,758 · 8K: 5,645 · 16K: 11,681 · 24K: 18,097 · 32K: 24,928
PIECEWISE parity-qualified partial · f01e/ac7509e2 2K: 1,363 · 4K: 2,475 · 16K: 10,545 · 24K: 16,544 · 32K: 22,672

evidence 1 · evidence 2 · evidence 3

Qwen3.8 AutoRound TP2 graph modes

Grade C · exact cache-zero HTTP serving · F16 KV · target-only MTP0 spans eager, current PIECEWISE partial, and dated FULL_AND_PIECEWISE profiles; native MTP1 and MTP2 each add one separately qualified current-PIECEWISE exact-4K point · all unplotted selectors remain missing

eager · f01e/ac7509e2PIECEWISE partial · f01e/ac7509e2MTP1 PIECEWISE · exact 4K Grade CMTP2 PIECEWISE · exact 4K Grade CFULL_AND_PIECEWISE · b2dd/1e90

eager · f01e/ac7509e2 2K: 9.65 · 4K: 10.04 · 8K: 10.11 · 16K: 10.12 · 24K: 10.15 · 32K: 10.2
PIECEWISE partial · f01e/ac7509e2 2K: 39.68 · 4K: 46.64 · 24K: 42.17 · 32K: 41.14
MTP1 PIECEWISE · exact 4K Grade C 4K: 13.74
MTP2 PIECEWISE · exact 4K Grade C 4K: 18.41
FULL_AND_PIECEWISE · b2dd/1e90 2K: 48.15 · 4K: 47.93 · 8K: 46.83 · 16K: 45.2 · 24K: 43.7 · 32K: 42.34

All measured metric summaries

Decode

TTFT

eager · f01e/ac7509e2 2K: 2,685 · 4K: 1,885 · 8K: 3,816 · 16K: 7,857 · 24K: 12,130 · 32K: 16,655
PIECEWISE partial · f01e/ac7509e2 2K: 1,703 · 4K: 1,563 · 24K: 10,230 · 32K: 14,062
MTP1 PIECEWISE · exact 4K Grade C 4K: 2,651
MTP2 PIECEWISE · exact 4K Grade C 4K: 3,126
FULL_AND_PIECEWISE · b2dd/1e90 2K: 782.76 · 4K: 1,569 · 8K: 3,187 · 16K: 6,611 · 24K: 10,265 · 32K: 14,105

evidence 1 · evidence 2 · evidence 3 · evidence 4 · evidence 5

Qwen3.8 AutoRound TP4 graph modes

Grade C · exact cache-zero HTTP serving · target-only MTP0 · F16 KV · current f01e eager spans 2K-32K, current f01e PIECEWISE plots target-identical 2K/4K/16K/24K/32K with 8K quarantined, and dated b2dd/1e90 FULL_AND_PIECEWISE remains a separate fully certified six-depth profile · x0 missing

eager expansion · f01e/ac7509e2eager 8K oracle · f01e/ac7509e2PIECEWISE partial · f01e/ac7509e2FULL_AND_PIECEWISE · b2dd/1e90

eager expansion · f01e/ac7509e2 2K: 9.83 · 4K: 10.05 · 16K: 10.26 · 24K: 10.17 · 32K: 10.21
eager 8K oracle · f01e/ac7509e2 8K: 9.65
PIECEWISE partial · f01e/ac7509e2 2K: 51.07 · 4K: 64.42 · 16K: 62.78 · 24K: 62.09 · 32K: 60.51
FULL_AND_PIECEWISE · b2dd/1e90 2K: 71.17 · 4K: 70.15 · 8K: 69.87 · 16K: 69.07 · 24K: 67.87 · 32K: 66.65

All measured metric summaries

Decode

TTFT

eager expansion · f01e/ac7509e2 2K: 2,686 · 4K: 1,421 · 16K: 5,845 · 24K: 8,965 · 32K: 12,210
eager 8K oracle · f01e/ac7509e2 8K: 4,768
PIECEWISE partial · f01e/ac7509e2 2K: 1,633 · 4K: 1,106 · 16K: 4,568 · 24K: 7,051 · 32K: 9,606
FULL_AND_PIECEWISE · b2dd/1e90 2K: 547.47 · 4K: 1,105 · 8K: 2,210 · 16K: 4,562 · 24K: 7,012 · 32K: 9,569

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.8 AutoRound TP1 eager MTP4 screened depth ◇ screened, experimental

Grade D lab-screened only · current f01e/ac7509e2 image · F16 KV · exact cache-zero target parity at 4K/16K/24K · 2K and cross-boot-conflict 8K quarantined · 32K runtime-closed · x=0 missing · zero lab-measured cells

MTP4 eager · human-adjudicated screened depths

MTP4 eager · human-adjudicated screened depths 4K: 14.85 · 16K: 12.36 · 24K: 13.12

All measured metric summaries

Decode

MTP4 eager · human-adjudicated screened depths 4K: 14.85 · 16K: 12.36 · 24K: 13.12

TTFT

MTP4 eager · human-adjudicated screened depths 4K: 2,999 · 16K: 12,356 · 24K: 19,104

Draft acceptance

MTP4 eager · human-adjudicated screened depths 4K: 0.61 · 16K: 0.56 · 24K: 0.64

evidence 1

Qwen3.8 AutoRound TP1 eager MTP3 partial depth

Grade D partial · current f01e/ac7509e2 image · F16 KV · exact target parity at 4K-32K · 2K quarantined at token 90 · x=0 missing · conventional 99-interval decode

MTP3 eager · exact-parity depths

MTP3 eager · exact-parity depths 4K: 13.25 · 8K: 14.12 · 16K: 11.3 · 24K: 11.23 · 32K: 11.17

All measured metric summaries

Decode

MTP3 eager · exact-parity depths 4K: 13.25 · 8K: 14.12 · 16K: 11.3 · 24K: 11.23 · 32K: 11.17

TTFT

MTP3 eager · exact-parity depths 4K: 2,970 · 8K: 5,970 · 16K: 12,285 · 24K: 18,970 · 32K: 26,063

Draft acceptance

MTP3 eager · exact-parity depths 4K: 0.7 · 8K: 0.78 · 16K: 0.61 · 24K: 0.66 · 32K: 0.61

evidence 1

Qwen3.8 AutoRound TP1 MTP2 depth

Current f01e/ac7509e2 image · F16 KV · eager Grade D partial plus a separate PIECEWISE exact-4K Grade C point with its original cache-gate omission disclosed and a passing post-run audit · 2K/8K/16K corruption caveats remain · conventional 99-interval decode

MTP2 eager · exact-parity depthsMTP2 PIECEWISE · exact 4K Grade C · cache defect disclosed

MTP2 eager · exact-parity depths 4K: 11.39 · 24K: 9.86 · 32K: 9.79
MTP2 PIECEWISE · exact 4K Grade C · cache defect disclosed 4K: 11.99

All measured metric summaries

Decode

MTP2 eager · exact-parity depths 4K: 11.39 · 24K: 9.86 · 32K: 9.79
MTP2 PIECEWISE · exact 4K Grade C · cache defect disclosed 4K: 11.99

TTFT

MTP2 eager · exact-parity depths 4K: 2,952 · 24K: 18,854 · 32K: 25,905
MTP2 PIECEWISE · exact 4K Grade C · cache defect disclosed 4K: 2,942

Draft acceptance

MTP2 eager · exact-parity depths 4K: 0.85 · 24K: 0.75 · 32K: 0.7
MTP2 PIECEWISE · exact 4K Grade C · cache defect disclosed 4K: 0.85

evidence 1 · evidence 2

Qwen3.8 AutoRound TP1 MTP1 depth

Current f01e/ac7509e2 image · eager/F16 Grade D partial plus separate Grade C exact-4K PIECEWISE/F16 and eager/E4M3 singletons · every unplotted depth remains missing; PIECEWISE 8K retains the token-99 corruption caveat · conventional 99-interval decode

MTP1 eager · exact-parity depthsMTP1 PIECEWISE/F16 · exact 4K Grade CMTP1 eager/E4M3 · exact 4K Grade C

MTP1 eager · exact-parity depths 4K: 8.31 · 16K: 7.8 · 24K: 7.62 · 32K: 7.72
MTP1 PIECEWISE/F16 · exact 4K Grade C 4K: 8.69
MTP1 eager/E4M3 · exact 4K Grade C 4K: 8.38

All measured metric summaries

Decode

MTP1 eager · exact-parity depths 4K: 8.31 · 16K: 7.8 · 24K: 7.62 · 32K: 7.72
MTP1 PIECEWISE/F16 · exact 4K Grade C 4K: 8.69
MTP1 eager/E4M3 · exact 4K Grade C 4K: 8.38

TTFT

MTP1 eager · exact-parity depths 4K: 2,937 · 16K: 12,146 · 24K: 18,746 · 32K: 25,748
MTP1 PIECEWISE/F16 · exact 4K Grade C 4K: 2,962
MTP1 eager/E4M3 · exact 4K Grade C 4K: 3,843

Draft acceptance

MTP1 eager · exact-parity depths 4K: 0.79 · 16K: 0.84 · 24K: 0.81 · 32K: 0.81
MTP1 PIECEWISE/F16 · exact 4K Grade C 4K: 0.79
MTP1 eager/E4M3 · exact 4K Grade C 4K: 0.94

evidence 1 · evidence 2 · evidence 3

Qwen3.8 AutoRound TP1/TP2/TP4 HTTP context

Grade C · exact cache-zero HTTP serving · current f01e eager TP2 MTP0/MTP1 and TP4 MTP0 span every nonzero depth; TP2 MTP2 and TP4 MTP1-3 have target-parity 4K-32K with 2K quarantined; TP4 prior 8K points are retained; TP4 MTP4 recovery measures 4K/16K/24K while 2K/8K remain quarantined and 32K is runtime-closed · F16 KV · x=0 missing

AutoRound INT4 · TP1 · F16 KVAutoRound INT4 · TP2 · F16 KVAutoRound INT4 · TP2 · MTP0 · F16 KV · current f01e eagerAutoRound INT4 · TP2 · MTP1 · F16 KV · current f01e eagerAutoRound INT4 · TP2 · MTP2 · F16 KV · current f01e eagerAutoRound INT4 · TP2 · MTP3 · F16 KV · current f01e eagerAutoRound INT4 · TP4 · F16 KV · graphAutoRound INT4 · TP4 · MTP0 · F16 KV · current f01e eager prior 8KAutoRound INT4 · TP4 · MTP0 · F16 KV · current f01e eager expansionAutoRound INT4 · TP4 · MTP0 · F16 KV · current f01e PIECEWISE partialAutoRound INT4 · TP4 · MTP1 · F16 KV · current f01e eagerAutoRound INT4 · TP4 · MTP1 · F16 KV · current f01e eager expansionAutoRound INT4 · TP4 · MTP2 · F16 KV · current f01e eagerAutoRound INT4 · TP4 · MTP2 · F16 KV · current f01e eager expansionAutoRound INT4 · TP4 · MTP3 · F16 KV · current f01e eagerAutoRound INT4 · TP4 · MTP3 · F16 KV · current f01e eager expansionAutoRound INT4 · TP4 · MTP4 · F16 KV · current f01e eager quality recovery

AutoRound INT4 · TP1 · F16 KV 2K: 30.1 · 4K: 29.77 · 8K: 29.28 · 16K: 28.42 · 24K: 27.66 · 32K: 26.99
AutoRound INT4 · TP2 · F16 KV 2K: 48.15 · 4K: 47.93 · 8K: 46.83 · 16K: 45.2 · 24K: 43.7 · 32K: 42.34
AutoRound INT4 · TP2 · MTP0 · F16 KV · current f01e eager 2K: 9.65 · 4K: 10.04 · 8K: 10.11 · 16K: 10.12 · 24K: 10.15 · 32K: 10.2
AutoRound INT4 · TP2 · MTP1 · F16 KV · current f01e eager 2K: 11.88 · 4K: 14.95 · 8K: 15.22 · 16K: 13.78 · 24K: 13.74 · 32K: 13.92
AutoRound INT4 · TP2 · MTP2 · F16 KV · current f01e eager 4K: 20.36 · 8K: 20.89 · 16K: 18.29 · 24K: 17.76 · 32K: 17.7
AutoRound INT4 · TP2 · MTP3 · F16 KV · current f01e eager 4K: 19.08 · 8K: 25.12 · 16K: 20.28 · 24K: 20.17 · 32K: 20.16
AutoRound INT4 · TP4 · F16 KV · graph 2K: 71.17 · 4K: 70.15 · 8K: 69.87 · 16K: 69.07 · 24K: 67.87 · 32K: 66.65
AutoRound INT4 · TP4 · MTP0 · F16 KV · current f01e eager prior 8K 8K: 9.65
AutoRound INT4 · TP4 · MTP0 · F16 KV · current f01e eager expansion 2K: 9.83 · 4K: 10.05 · 16K: 10.26 · 24K: 10.17 · 32K: 10.21
AutoRound INT4 · TP4 · MTP0 · F16 KV · current f01e PIECEWISE partial 2K: 51.07 · 4K: 64.42 · 16K: 62.78 · 24K: 62.09 · 32K: 60.51
AutoRound INT4 · TP4 · MTP1 · F16 KV · current f01e eager 8K: 13.71
AutoRound INT4 · TP4 · MTP1 · F16 KV · current f01e eager expansion 4K: 15.96 · 16K: 15.12 · 24K: 14.88 · 32K: 14.77
AutoRound INT4 · TP4 · MTP2 · F16 KV · current f01e eager 8K: 18.08
AutoRound INT4 · TP4 · MTP2 · F16 KV · current f01e eager expansion 4K: 21.84 · 16K: 19.74 · 24K: 19.44 · 32K: 19.4
AutoRound INT4 · TP4 · MTP3 · F16 KV · current f01e eager 8K: 21.08
AutoRound INT4 · TP4 · MTP3 · F16 KV · current f01e eager expansion 4K: 25.32 · 16K: 21.94 · 24K: 22.09 · 32K: 22.55
AutoRound INT4 · TP4 · MTP4 · F16 KV · current f01e eager quality recovery 4K: 21.97 · 16K: 23.79 · 24K: 25.75

evidence 1 · evidence 2 · evidence 3 · evidence 4 · evidence 5 · evidence 6 · evidence 7 · evidence 8 · evidence 9 · evidence 10 · evidence 11 · evidence 12 · evidence 13 · evidence 14 · evidence 15 · evidence 16 · evidence 17

Q5_K_S external-MTP 8K route screen ◇ screened, experimental

Grade C diagnostic · TP1 · graph off · Q8_0 target/draft KV · MTP1–4 all diverge from MTP0 at token index 6; rates and draft counters have no certified or headline authority

exact-8K screened routes

exact-8K screened routes 0: 17.5 · 1: 21.75 · 2: 24.99 · 3: 24.29 · 4: 22.85

All measured metric summaries

Decode

exact-8K screened routes 0: 17.5 · 1: 21.75 · 2: 24.99 · 3: 24.29 · 4: 22.85

Draft acceptance

exact-8K screened routes 1: 0.75 · 2: 0.62 · 3: 0.55 · 4: 0.48

evidence 1

Qwen3.6 Q8 target-only MTP0 SYCL graph

TP1 · raw-engine pp2048/tg128 · decode fully graph-certified; prefill mixed partial above depth 0 · both target-only and embedded-artifact graph lanes are slower than their matched graph-off controls and are evidence, not optimization wins

F16 KV · SYCL graphQ8_0 KV · SYCL graphEmbedded-MTP artifact · F16 KV · SYCL graph (MTP0)Embedded-MTP artifact · Q8_0 KV · SYCL graph (MTP0)

F16 KV · SYCL graph 0: 19.36 · 2K: 19.18 · 4K: 19.08 · 8K: 18.85 · 16K: 18.41 · 24K: 18.01 · 32K: 17.62
Q8_0 KV · SYCL graph 0: 19.01 · 2K: 17.93 · 4K: 17.03 · 8K: 15.47 · 16K: 12.71 · 24K: 10.84 · 32K: 9.48
Embedded-MTP artifact · F16 KV · SYCL graph (MTP0) 0: 19.36 · 2K: 19.18 · 4K: 19.08 · 8K: 18.85 · 16K: 18.41 · 24K: 18.01 · 32K: 17.62
Embedded-MTP artifact · Q8_0 KV · SYCL graph (MTP0) 0: 19.02 · 2K: 17.94 · 4K: 17.04 · 8K: 15.47 · 16K: 12.71 · 24K: 10.83 · 32K: 9.48

All measured metric summaries

Decode

Prefill

F16 KV · SYCL graph 0: 899.27 · 2K: 881.87 · 4K: 855.86 · 8K: 818.75 · 16K: 753.44 · 24K: 698.23 · 32K: 651.49
Q8_0 KV · SYCL graph 0: 896.09 · 2K: 871.93 · 4K: 850.23 · 8K: 809.77 · 16K: 747.45 · 24K: 691.4 · 32K: 645.09
Embedded-MTP artifact · F16 KV · SYCL graph (MTP0) 0: 900.48 · 2K: 880.57 · 4K: 857.1 · 8K: 819.59 · 16K: 755.79 · 24K: 700.16 · 32K: 650.65
Embedded-MTP artifact · Q8_0 KV · SYCL graph (MTP0) 0: 894.19 · 2K: 871.09 · 4K: 849.46 · 8K: 810.19 · 16K: 747.51 · 24K: 692.95 · 32K: 644.91

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.6 Q4_K_M KV × graph comparison

TP1 · target-only MTP0 · raw-engine pp2048/tg128 · F16 and Q8_0 graph decode are fully certified while prefill is mixed partial above depth 0 · graph-on is slower in both phases at all seven depths for both KV types, so the protected graph-off lanes remain recommended

F16 KV · graph off · protected controlF16 KV · SYCL graph · quality-qualified negativeQ8_0 KV · graph off · protected controlQ8_0 KV · SYCL graph · quality-qualified negative

F16 KV · graph off · protected control 0: 29.3 · 2K: 28.72 · 4K: 28.42 · 8K: 27.88 · 16K: 26.88 · 24K: 25.9 · 32K: 25.04
F16 KV · SYCL graph · quality-qualified negative 0: 28.58 · 2K: 28.05 · 4K: 27.66 · 8K: 27.18 · 16K: 26.23 · 24K: 25.37 · 32K: 24.57
Q8_0 KV · graph off · protected control 0: 28.43 · 2K: 25.91 · 4K: 24 · 8K: 20.92 · 16K: 16.17 · 24K: 13.3 · 32K: 11.38
Q8_0 KV · SYCL graph · quality-qualified negative 0: 27.77 · 2K: 25.39 · 4K: 23.53 · 8K: 20.56 · 16K: 16.01 · 24K: 13.21 · 32K: 11.3

All measured metric summaries

Decode

Prefill

F16 KV · graph off · protected control 0: 826.36 · 2K: 891.6 · 4K: 869.41 · 8K: 830.46 · 16K: 760.02 · 24K: 704.98 · 32K: 655.1
F16 KV · SYCL graph · quality-qualified negative 0: 818.2 · 2K: 800.16 · 4K: 780.49 · 8K: 746.47 · 16K: 688.17 · 24K: 644.83 · 32K: 606.64
Q8_0 KV · graph off · protected control 0: 824.62 · 2K: 884.56 · 4K: 865.76 · 8K: 823.1 · 16K: 756.11 · 24K: 698.56 · 32K: 648.26
Q8_0 KV · SYCL graph · quality-qualified negative 0: 815.88 · 2K: 795.43 · 4K: 775.66 · 8K: 738.53 · 16K: 684.94 · 24K: 639.66 · 32K: 600.52

evidence 1 · evidence 2 · evidence 3 · evidence 4

Qwen3.6 embedded-Q8 MTP serving context

TP1 · F16 KV · graph off · HTTP serving · sealed MTP0/MTP3 R3 plus quality-R1 MTP1/2/4 · exact target-output parity · x=0 means zero prior context plus one prompt token

MTP0 · sealed control profileMTP1 · exact-output speculativeMTP2 · exact-output speculativeMTP3 · exact-output speculativeMTP4 · exact-output speculative

MTP0 · sealed control profile 0: 17.19 · 2K: 16.85 · 4K: 16.61 · 8K: 16.1 · 16K: 15.18 · 24K: 14.36 · 32K: 13.63
MTP1 · exact-output speculative 0: 28.48 · 2K: 27.55 · 4K: 27.73 · 8K: 27.57 · 16K: 24.9 · 24K: 24.74 · 32K: 23.73
MTP2 · exact-output speculative 0: 37.24 · 2K: 32.32 · 4K: 36.23 · 8K: 36.32 · 16K: 29.04 · 24K: 30.58 · 32K: 29.2
MTP3 · exact-output speculative 0: 39.51 · 2K: 34.66 · 4K: 39.73 · 8K: 41.39 · 16K: 31.84 · 24K: 34.25 · 32K: 31.81
MTP4 · exact-output speculative 0: 37.56 · 2K: 34.69 · 4K: 41.65 · 8K: 43.72 · 16K: 31.41 · 24K: 36.49 · 32K: 31.28

evidence 1 · evidence 2

Qwen3.6 embedded-Q8 MTP Q8_0-KV screen ◇ screened, experimental

Grade C exploratory · TP1 · graph off · Q8_0 KV · deterministic 2K route divergence

MTP1 · Grade C screenedMTP2 · Grade C screenedMTP3 · Grade C screenedMTP4 · Grade C screened

MTP1 · Grade C screened 0: 28.59 · 2K: 25.74 · 4K: 24.75 · 8K: 22.66 · 16K: 17.99 · 24K: 16.15 · 32K: 13.53
MTP2 · Grade C screened 0: 31.98 · 2K: 31.56 · 4K: 35.08 · 8K: 34.49 · 16K: 26.72 · 24K: 30.42 · 32K: 25.46
MTP3 · Grade C screened 0: 33.04 · 2K: 32.9 · 4K: 38.34 · 8K: 39.29 · 16K: 27.53 · 24K: 35.84 · 32K: 27.32
MTP4 · Grade C screened 0: 32.53 · 2K: 31.8 · 4K: 40.28 · 8K: 39.87 · 16K: 27.68 · 24K: 39.25 · 32K: 26.7

All measured metric summaries

Decode

Draft acceptance

MTP1 · Grade C screened 0: 0.91 · 2K: 0.84 · 4K: 0.9 · 8K: 0.92 · 16K: 0.85 · 24K: 0.97 · 32K: 0.9
MTP2 · Grade C screened 0: 0.88 · 2K: 0.66 · 4K: 0.85 · 8K: 0.87 · 16K: 0.7 · 24K: 0.94 · 32K: 0.81
MTP3 · Grade C screened 0: 0.73 · 2K: 0.54 · 4K: 0.77 · 8K: 0.84 · 16K: 0.59 · 24K: 0.93 · 32K: 0.73
MTP4 · Grade C screened 0: 0.59 · 2K: 0.45 · 4K: 0.69 · 8K: 0.75 · 16K: 0.52 · 24K: 0.92 · 32K: 0.63

evidence 1

Qwen3.6 Q8 (no draft) as the conversation grows

TP1 llama.cpp SYCL · discrete evidence, not a fitted curve · baseline x is one exact prompt count; each VDR2 x is the larger anchor of a disclosed two-prompt median with band-specific ubatch

DNN-off cold retrievalVDR2 exact full-512 bands

DNN-off cold retrieval 4.4K: 15.24 · 17.3K: 14.03 · 31.8K: 12.78
VDR2 exact full-512 bands 4.4K: 16.59 · 17.3K: 15.14 · 31.8K: 13.69

All measured metric summaries

Decode

DNN-off cold retrieval 4.4K: 15.24 · 17.3K: 14.03 · 31.8K: 12.78
VDR2 exact full-512 bands 4.4K: 16.59 · 17.3K: 15.14 · 31.8K: 13.69

Prefill

VDR2 exact full-512 bands 4.4K: 606.07 · 17.3K: 157.7 · 31.8K: 628.79

TTFT

VDR2 exact full-512 bands 4.4K: 7,187 · 17.3K: 109,246 · 31.8K: 50,691

evidence 1 · evidence 2

Qwen3.6 Q8 with draft depth 3

TP1 llama.cpp SYCL · same embedded publisher-MTP artifact · fixed cold 12-prompt realistic suite · full token/content exact against the fresh control

matched artifact control → MTP3 · draft tokens 017.11tok/sDecodeconventional 99 interval fixed cold realistic · matched fresh control · evidence

matched artifact control → MTP3 · draft tokens 336.05tok/sDecode0.58ratioDraft acceptance2.72tokensEffective tokens / verifyconventional 99 interval fixed cold realistic · localmaxxing approved research scope not production · evidence

matched artifact control → MTP3

evidence 1

Qwen3.6 Q8 at long context

TP1 llama.cpp VDR2 · exact control/MTP3 full outputs · two-wave functional screen, not an isolated headline

target only · active context tokens 17.3K15.02tok/sDecodeparallel functional full 512 depth screen · evidence

target only · active context tokens 31.8K13.66tok/sDecodeparallel functional full 512 depth screen · evidence

integrated MTP3 · active context tokens 17.3K44.49tok/sDecode0.97ratioDraft acceptanceparallel functional full 512 depth screen · evidence

integrated MTP3 · active context tokens 31.8K41.49tok/sDecode0.97ratioDraft acceptanceparallel functional full 512 depth screen · evidence

target onlyintegrated MTP3

evidence 1

Fine print

Qwen3_5ForConditionalGeneration, 64 layers, 5120 hidden. architecture-matched; individual runtime paths require revision-specific exercise. Measurements, artifact hashes, outputs, quality decisions, and speed stay pinned to their exact recorded identity.

9 combinations the lab has stopped pursuing: The fixed cache-zero MTP5/7/9 policy screen peaked at 31.165 conventional tok/s and remained noncompetitive with the separately qualified vLLM AutoRound route. Reopen only for a material backend, kernel, or quantization change—not another generic flag sweep. evidence The intensive native-DFlash program ended at a confirmed 47.341 conventional tok/s on one B70, below its 100 tok/s TP1 objective. The patch and safety evidence remain preserved; reopen only for a new mechanism, not generic configuration churn. evidence The measured c1 MTP3 footprint was 29,911 MiB and projected c2 was 32,683 MiB, leaving no useful safe headroom. The c2 launch was not performed and CPU offload was not used. evidence The model has 16 GDN key heads, which cannot be evenly partitioned across TP3. evidence The XPU FlashAttention backend rejects fp8_e5m2 KV cache at engine initialization. evidence The current f01e/ac7509e2 engine explicitly rejects fp8_e5m2 KV at initialization; startup, canary, and quality did not run, so this closure does not transfer to graph-on, other TP/MTP, KV formats, or runtimes. evidence The least-compounded TP1 target-only anchor matched only 3/20 stable-set oracle outputs. evidence The exact 8K graph/Q8_0-KV request passed parity, but the required long-context quality case aborted at qptr->wait during command-graph recording. No speed is published; sibling depths remain missing and require their own evidence. evidence The checksum-pinned target-only Q8_0 file has mtp_included=false. Same-artifact intrinsic MTP1–4 therefore requires a different MTP-bearing artifact/package and is excluded from the target-only contract; this does not close separately bound external drafts. evidence Exact selectors are in the family data.

Many people at once Lab-measured

16 simultaneous users share 68 combined tok/s on 1 card (Q4_K_M target with external Q4_0 MTP2 · 16-slot/8K-total HTTP); the full curve is under Measured results and in the multi-user report.

How much faster could this get? Projected — not measured

The ML Bottleneck physics engine projects a tuned-run target and the physical ceiling for the headline setup (qwen3.8-27b · AutoRound INT4 · vLLM XPU nightly + TP4 historical-winner overlay · TP4 · MTP0 · graph on · f16 KV · max context 32,768 · GPU memory 0.6 · natural EOS). The grade is optimization headroom against the tuned-run target, not model quality.