Repository navigation

fix: correct TRT-LLM MLA labels and GPT-OSS MXFP4 profiles - #264

Merged

simone-chen merged 29 commits into

mainfrom

simonec/trt-gym-missing-profiles

Sep 20, 2026

Merged

fix: correct TRT-LLM MLA labels and GPT-OSS MXFP4 profiles#264
simone-chen merged 29 commits into
mainfrom
simonec/trt-gym-missing-profiles

Commits on Sep 17, 2026
- fix: record TRT MLA precision and allow native MXFP4 padding
  simone-chen
  committed
  ce9a396
Commits on Sep 18, 2026
- fix: refresh TRT profiles and recover all 263 Gym predictions
  simone-chen
  committed
  b73705b
- fix: align GPT-OSS MoE bias and native autotuning
  simone-chen
  committed
  3830b5a
- fix: refresh tuned GPT-OSS profiles and verify Gym TPOT
  simone-chen
  committed
  a85b9e1
- docs: verify GPT-OSS serving activation precision mismatch
  simone-chen
  committed
  9076236
- fix: align TRT W4A8 data and replay request sampling
  simone-chen
  committed
  6b79d38
- Merge main and preserve replay sampler documentation
  simone-chen
  committed
  5afc9d6
- ci: fetch numerical baseline outside squashed main history
  simone-chen
  committed
  1c3d30b
- fix: align collection provenance and validate NumPy seeds
  simone-chen
  committed
  e36bfe5
- docs: record matched TRT residual diagnosis
  simone-chen
  committed
  8bcc5d1
- docs: trace GPT-OSS eager execution gap
  simone-chen
  committed
  0c25eb9
- docs: clarify Qwen scheduling evidence scope
  simone-chen
  committed
  cee6fdd
- chore: merge main and reuse baseline fetch helper
  simone-chen
  committed
  7fc50d0
- fix: bound length sampler to direct synthetic replay
  simone-chen
  committed
  7d0bcac
- fix: reject length sampler at workload driver boundary
  simone-chen
  committed
  8ce3ef3
- chore: merge SGLang profile and CI fixes from main
  simone-chen
  committed
  3bcbb43
- fix: validate sampler before selecting replay execution mode
  simone-chen
  committed
  5633a8f
- fix: encode unavailable MoE power and align profile integrity checks
  simone-chen
  committed
  a70df0a
- Merge main package consolidation into TRT profile refresh
  simone-chen
  committed
  b2193a4
- test: verify independent seeded replays
  simone-chen
  committed
  5518ec4
- test: cover TRT collector boundaries and clarify evidence
  simone-chen
  committed
  3c7edb1
- Merge main documentation updates into TRT profile refresh
  simone-chen
  committed
  f50d238
- Merge context-limit fix into TRT profile refresh
  simone-chen
  committed
  f26a1bc
- Merge replay correctness fixes into TRT profile refresh
  simone-chen
  committed
  f8c47a9
- fix: reject MoE collection when all autotuning attempts fail
  simone-chen
  committed
  076ca3a
- Merge FPM DCP identity fix into TRT profile refresh
  simone-chen
  committed
  8fa73a0
Commits on Sep 19, 2026

Navigation Menu

Repository navigation

fix: correct TRT-LLM MLA labels and GPT-OSS MXFP4 profiles - #264

fix: correct TRT-LLM MLA labels and GPT-OSS MXFP4 profiles#264
simone-chen merged 29 commits into
mainfrom
simonec/trt-gym-missing-profiles

Commits on Sep 17, 2026

fix: record TRT MLA precision and allow native MXFP4 padding

Commits on Sep 18, 2026

fix: refresh TRT profiles and recover all 263 Gym predictions

fix: align GPT-OSS MoE bias and native autotuning

fix: refresh tuned GPT-OSS profiles and verify Gym TPOT

docs: verify GPT-OSS serving activation precision mismatch

fix: align TRT W4A8 data and replay request sampling

Merge main and preserve replay sampler documentation

ci: fetch numerical baseline outside squashed main history

fix: align collection provenance and validate NumPy seeds

docs: record matched TRT residual diagnosis

docs: trace GPT-OSS eager execution gap

docs: clarify Qwen scheduling evidence scope

chore: merge main and reuse baseline fetch helper

fix: bound length sampler to direct synthetic replay

fix: reject length sampler at workload driver boundary

chore: merge SGLang profile and CI fixes from main

fix: validate sampler before selecting replay execution mode

fix: encode unavailable MoE power and align profile integrity checks

Merge main package consolidation into TRT profile refresh

test: verify independent seeded replays

test: cover TRT collector boundaries and clarify evidence

Merge main documentation updates into TRT profile refresh

Merge context-limit fix into TRT profile refresh

Merge replay correctness fixes into TRT profile refresh

fix: reject MoE collection when all autotuning attempts fail

Merge FPM DCP identity fix into TRT profile refresh

Commits on Sep 19, 2026

Merge DeepSeek V4.1 support into TRT profile refresh

docs: clarify non-DP readmission evidence scope

Merge current replay and Sweeper changes into TRT profile refresh