Repository navigation

DeepSpeed parity hacks - #502

Closed

bigximik wants to merge 25 commits into

mainfrom

gspo

Commits on Apr 27, 2026
- grpo: add policy-gradient metrics behind compute_extra_metrics flag
  bigximik
  committed
  98836e6
- grpo: align metric names with DeepSpeed path
  bigximik
  committed
  b856e39
Commits on Apr 28, 2026
Commits on Apr 29, 2026
- grpo: temperature scaling for IS ratio parity with actor sampling
  bigximik
  committed
  014ba59
Commits on May 4, 2026
- head: fp32_lm_head flag to match vLLM bf16_last_layer_fp32 precision
  bigximik
  committed
  d8cb9ef
- head: fix fp32_lm_head gradient flow via detach + manual weight grad accumulation
  bigximik
  committed
  0f90f20
Commits on May 5, 2026
- grpo: decouple loss/gradient divisors and fix SDP loss double-counting
  bigximik
  committed
  557a3c4
- Merge branch 'main' into grpo-metrics
  jlamypoirier
  authored
  a1b3f32
- Merge branch 'grpo-metrics' into gspo
  jlamypoirier
  authored
  9c07626
- grpo: address review feedback on metrics
  
  jlamypoirier
  and
  claude
  committed
  d360a46
- grpo: address review follow-ups
  
  jlamypoirier
  and
  claude
  committed
  bb6315c
- grpo: round-3 review fixes
  
  jlamypoirier
  and
  claude
  committed
  89ed062
- grpo: GRPOMetrics as NamedTuple
  
  jlamypoirier
  and
  claude
  committed
  b0852fd
- grpo: fix entropy under tensor-parallel + minor review fixes
  
  jlamypoirier
  and
  claude
  committed
  61ad4f7
- Merge remote-tracking branch 'origin/grpo-metrics' into gspo
  jlamypoirier
  committed
  15ae8d6
Commits on May 6, 2026
- Merge remote-tracking branch 'origin/main' into gspo
  jlamypoirier
  committed
  dfd2ce3
- Merge remote-tracking branch 'origin/main' into gspo
  jlamypoirier
  committed
  2fc2dfe
- gspo: address coarse-review easy items (#7, #13, #14)
  
  jlamypoirier
  and
  claude
  committed
  8547a56
Commits on May 7, 2026
- gspo: register as a sibling loss type instead of policy_loss switch
  
  jlamypoirier
  and
  claude
  committed
  fc96d07
- gspo: collapse loss subclasses, dispatch kernel via self._call_kernel
  
  jlamypoirier
  and
  claude
  committed
  d2c051a
- gspo: dispatch via self._forward = <kernel> instead of wrapper method
  
  jlamypoirier
  and
  claude
  committed
  0d0185c
Commits on May 15, 2026
- Merge remote-tracking branch 'origin/main' into gspo
  jlamypoirier
  committed
  ffeebd9
- gspo: drop the temperature field, use base logits_scale_factor
  
  jlamypoirier
  and
  claude
  committed
  227f8bc

Navigation Menu

Repository navigation

DeepSpeed parity hacks - #502

DeepSpeed parity hacks#502
bigximik wants to merge 25 commits into
mainfrom
gspo

Commits on Apr 27, 2026

grpo: add policy-gradient metrics behind compute_extra_metrics flag

grpo: align metric names with DeepSpeed path

Commits on Apr 28, 2026

gspo: add sequence-level IS-ratio clipping loss

schedule: add rollouts_per_step to auto-set depth_first_micro_batches

grpo: dynamic docs_per_step accumulation and normalize_by_documents

Commits on Apr 29, 2026

grpo: temperature scaling for IS ratio parity with actor sampling

Commits on May 4, 2026

head: fp32_lm_head flag to match vLLM bf16_last_layer_fp32 precision

head: fix fp32_lm_head gradient flow via detach + manual weight grad accumulation

Commits on May 5, 2026

grpo: decouple loss/gradient divisors and fix SDP loss double-counting

Merge branch 'main' into grpo-metrics

Merge branch 'grpo-metrics' into gspo

grpo: address review feedback on metrics

grpo: address review follow-ups

grpo: round-3 review fixes

grpo: GRPOMetrics as NamedTuple

grpo: fix entropy under tensor-parallel + minor review fixes

Merge remote-tracking branch 'origin/grpo-metrics' into gspo

Commits on May 6, 2026

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

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

gspo: address coarse-review easy items (#7, #13, #14)

Commits on May 7, 2026

gspo: register as a sibling loss type instead of policy_loss switch

gspo: collapse loss subclasses, dispatch kernel via self._call_kernel

gspo: dispatch via self._forward = <kernel> instead of wrapper method

Commits on May 15, 2026

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

gspo: drop the temperature field, use base logits_scale_factor