Repository navigation

test(nightly): distributed SFT functionality suite - #4077

Draft

garrett361 wants to merge 41 commits into

mainfrom

feat/nightly-distributed-tests

Draft

test(nightly): distributed SFT functionality suite#4077
garrett361 wants to merge 41 commits into
mainfrom
feat/nightly-distributed-tests

Commits on Nov 4, 2026
- test(nightly): add configs and data for the distributed SFT suite
  
  garrett361
  and
  claude
  committed
  f94a367
- test(nightly): add a distributed SFT functionality suite
  
  garrett361
  and
  claude
  committed
  a031f4f
- test(nightly): run every layout per arch, excluding cells only via skip
  
  garrett361
  and
  claude
  committed
  785c1d2
- fix(nightly): give the tiny GLM-5 config 32 attention heads
  
  garrett361
  and
  claude
  committed
  fda66dd
- test(nightly): name each feature row so arch skips can reference it
  
  garrett361
  and
  claude
  committed
  9263bfb
- test(nightly): drop the JUnit XML report from the distributed launcher
  
  garrett361
  and
  claude
  committed
  714ea74
- test(nightly): keep each arch's tokenizer next to its tiny config
  
  garrett361
  and
  claude
  committed
  748b06b
- docs(nightly): explain the SFT case matrix and -k selection in cases.py
  
  garrett361
  and
  claude
  committed
  f3a9371
- test(nightly): cover every registered architecture in the distributed SFT suite
  
  garrett361
  and
  claude
  committed
  d054545
- fix(nightly): adapt the SFT matrix to Ulysses-only context parallelism
  
  garrett361
  and
  claude
  committed
  c3e3b66
- refactor(tests): move the distributed suite and parallel invariance test under tests/regression/distributed
  
  garrett361
  and
  claude
  committed
  1d101eb
- test(regression): retarget the distributed SFT suite to 4-GPU runners
  
  garrett361
  and
  claude
  committed
  c3689ec
- test(regression): run llama through the ungated unsloth tokenizer mirror
  
  garrett361
  and
  claude
  committed
  8f3bcca
- test(regression): give the tiny GLM-5 and DSv4 configs FlashMLA-compatible shapes
  
  garrett361
  and
  claude
  committed
  b238808
- test(regression): replace the router-dtype row with a fullgraph=True row
  
  garrett361
  and
  claude
  committed
  dd39b1f
- test(regression): build tiny models from the published HF configs, resume on qwen3_moe only
  
  garrett361
  and
  claude
  committed
  133afd4
- fix(regression): leave the tied lm_head out of tiny checkpoints
  
  garrett361
  and
  claude
  committed
  c7b2bc2
- test(regression): fail cp2 cases on any graph break, reported apart from training
  
  garrett361
  and
  claude
  committed
  4f717e3
- test(regression): narrow the tiny MoE checkpoints
  
  garrett361
  and
  claude
  committed
  9220383
- test(regression): cut the SFT matrix, shard it and harden its launches
  
  garrett361
  and
  claude
  committed
  c0e7d83
- test(regression): drop the eager, AC-off and unfused feature rows
  
  garrett361
  and
  claude
  committed
  add08a0
- refactor(regression): accumulate shard costs in a defaultdict
  
  garrett361
  and
  claude
  committed
  e9c092d
- docs(regression): explain why only qwen3_moe runs feature rows and resume
  
  garrett361
  and
  claude
  committed
  8f0d7c6
- refactor(regression): build an arch's cases with Arch.cases()
  
  garrett361
  and
  claude
  committed
  947efe8
- refactor(regression): define archs and features inline, derive -k names from case ids
  
  garrett361
  and
  claude
  committed
  27f1e06
- refactor(regression): define each arch once, with its tiny config and tokenizer
  
  garrett361
  and
  claude
  committed
  5c76dba
- docs(regression): add a README for the distributed SFT suite
  
  garrett361
  and
  claude
  committed
  0e9b5e1
- refactor(regression): drop the edge-case guards from SFT runs
  
  garrett361
  and
  claude
  committed
  dd0f610
- refactor(regression): split the collection hook into named shard and GPU steps
  
  garrett361
  and
  claude
  committed
  daa757a
- refactor(regression): keep the matrix in archs.py and the run helpers in harness.py
  
  garrett361
  and
  claude
  committed
  89bd77b
- fix(regression): keep failures in the suite summary, give runs 900 s
  
  garrett361
  and
  claude
  committed
  a0fd105
- refactor(regression): move the tiny models into a shared tiny_models.py
  
  garrett361
  and
  claude
  committed
  d43cca7
- fix(regression): delete the run dir before the train run
  
  garrett361
  and
  claude
  committed
  e18bb96
- test(regression): add a DeepEP feature row on qwen3_moe
  
  garrett361
  and
  claude
  committed
  02d109f
- test(regression): make GLM-5 and DSv4 sparse attention drop keys
  
  garrett361
  and
  claude
  committed
  f5ba143
- test(regression): check that a resumed run reproduces the original losses
  
  garrett361
  and
  claude
  committed
  95a0bc8
- test(regression): turn off the prime monitor in suite runs
  
  garrett361
  and
  claude
  committed
  48fa885
- test(regression): drop SFT data samples with personal details
  
  garrett361
  and
  claude
  committed
  e2ac888
- test(regression): drop SFT data samples that credit named people
  
  garrett361
  and
  claude
  committed
  532ca4e
- test(regression): drop four more SFT data samples with named people
  
  garrett361
  and
  claude
  committed
  5432baa
- test(regression): replace the SFT chat data with truncated math samples
  
  garrett361
  and
  claude
  committed
  1aac3c7

Commits

Commits on Nov 4, 2026