Repository navigation

Add fp32 LM head, docs_per_step accumulation, and document-count metrics - #520

Open

jlamypoirier wants to merge 18 commits into

mainfrom

jlp_rl_features

Commits on May 27, 2026
- Add fp32_lm_head flag for vLLM precision parity
  
  jlamypoirier
  and
  claude
  committed
  e26ba9f
- Add docs_per_step for dynamic microbatch accumulation
  
  jlamypoirier
  and
  claude
  committed
  5afa8c7
Commits on Jun 3, 2026
- Merge branch 'main' into jlp_fp32_lm_head
  jlamypoirier
  authored
  6b35817
- Merge branch 'jlp_fp32_lm_head' into jlp_rl_features
  jlamypoirier
  authored
  46d30ca
Commits on Jun 25, 2026
- Log num_documents and documents_seen training metrics
  
  jlamypoirier
  and
  claude
  committed
  cc8768e
- Merge remote-tracking branch 'origin/main' into jlp_fp32_lm_head
  jlamypoirier
  committed
  31c3376
- Merge branch 'jlp_fp32_lm_head' into jlp_rl_features
  jlamypoirier
  committed
  a6028e7
- Fix GSPO segment index out-of-bounds on padded sequences
  
  jlamypoirier
  and
  claude
  committed
  b57b729
Commits on Jun 30, 2026
- Merge remote-tracking branch 'origin/main' into HEAD
  jlamypoirier
  committed
  2618def
Commits on Jul 10, 2026
- Merge branch 'main' into jlp_rl_features
  
  jlamypoirier
  and
  claude
  committed
  5a7dbaa
- Document GSPO/GRPO RL recipe and reference Qwen configs
  
  jlamypoirier
  and
  claude
  committed
  6c7c689
- RL recipe: GSPO metrics available (#555), add lag/instability measurement
  
  jlamypoirier
  and
  claude
  committed
  d2f4f23
Commits on Jul 13, 2026
- RL recipe: lower ZeRO stage (0.5B→1, 7B→2), document DeepSpeed-side stage
  
  jlamypoirier
  and
  claude
  committed
  2557e71
- RL recipe: mark example YAMLs as illustrative, point to the operative config
  
  jlamypoirier
  and
  claude
  committed
  d3e3186
- RL recipe: operative config is per-backend (shared base + fllm/ds), not one file
  
  jlamypoirier
  and
  claude
  committed
  20de9cc
- docs(rl): 7B ZeRO stage 3 (recompute leaves memory tight for stage 2)
  
  jlamypoirier
  and
  claude
  committed
  36dd34a
Commits on Jul 15, 2026
- Merge remote-tracking branch 'origin/main' into jlp_rl_features
  jlamypoirier
  committed
  73e4ba6
- Measure RL data-wait in the prefetch loop, not the schedule
  
  jlamypoirier
  and
  claude
  committed
  d3d2022

Navigation Menu

Repository navigation

Add fp32 LM head, docs_per_step accumulation, and document-count metrics - #520

Add fp32 LM head, docs_per_step accumulation, and document-count metrics#520
jlamypoirier wants to merge 18 commits into
mainfrom
jlp_rl_features

Commits on May 27, 2026

Add fp32_lm_head flag for vLLM precision parity

Add docs_per_step for dynamic microbatch accumulation

Commits on Jun 3, 2026

Merge branch 'main' into jlp_fp32_lm_head

Merge branch 'jlp_fp32_lm_head' into jlp_rl_features

Commits on Jun 25, 2026

Log num_documents and documents_seen training metrics

Merge remote-tracking branch 'origin/main' into jlp_fp32_lm_head

Merge branch 'jlp_fp32_lm_head' into jlp_rl_features

Fix GSPO segment index out-of-bounds on padded sequences

Commits on Jun 30, 2026

Merge remote-tracking branch 'origin/main' into HEAD

Commits on Jul 10, 2026

Merge branch 'main' into jlp_rl_features

Document GSPO/GRPO RL recipe and reference Qwen configs

RL recipe: GSPO metrics available (#555), add lag/instability measurement

Commits on Jul 13, 2026

RL recipe: lower ZeRO stage (0.5B→1, 7B→2), document DeepSpeed-side stage

RL recipe: mark example YAMLs as illustrative, point to the operative config

RL recipe: operative config is per-backend (shared base + fllm/ds), not one file

docs(rl): 7B ZeRO stage 3 (recompute leaves memory tight for stage 2)

Commits on Jul 15, 2026

Merge remote-tracking branch 'origin/main' into jlp_rl_features

Measure RL data-wait in the prefetch loop, not the schedule