deepspeedai/DeepSpeed

Pull Requests Commits

move trainer to deepspeed/runtime/rlhf/trainer

delock committed 86 days ago

aa015931

fix: DynamicCache compat for transformers>=5 in _cb_replace_slot

delock committed 86 days ago

9a943f2d

restore forward hooks after rollout

delock committed 88 days ago

7b352113

set capture mode to global

delock committed 88 days ago

7c8bbb62

fix: enable CUDA graph capture with DeepSpeed hybrid engine

delock committed 91 days ago

2d34fd3e

refactor(rollout): graph capture CB uses round-based reset instead of oversized cache

delock committed 91 days ago

9b4d823e

fix(rollout): graph capture CB - allocate sufficient cache for multi-round decode

delock committed 91 days ago

bfac5063

Rename student_engine -> engine, sync_weights_from_student -> sync_weights

delock committed 92 days ago

e816c457

Add kv_trim_threshold parameter to HybridEngineRollout

delock committed 92 days ago

739079b5

Move rollout engine to deepspeed/runtime/rollout/ as reusable module

delock committed 92 days ago

1abb504d

Add continuous batching with shared prefix, KV trim, and early exit to OPSD rollout engine

delock committed 92 days ago

a897889e

Add OPSD vLLM rollout scaffold, Qwen2/Qwen3 weight bridges, and README

PKUWZP committed 94 days ago

6384396b

Add OPSD trainer, hybrid-engine rollout, and end-to-end entry point

PKUWZP committed 94 days ago

837787a0

Add OPSD frozen teacher with CPU logit cache + tests

PKUWZP committed 94 days ago

14d8fe7e

Add OPSD example: config, divergence losses, utils + tests

PKUWZP committed 94 days ago

932f0b52

Respect `$TRITON_HOME` (#7907)

Flamefire committed 158 days ago

Verified 38bd11a7

Update version (#7903)

loadams committed 168 days ago

Verified 5f7b6870

fix(zero3): use current_stream() instead of default_stream() for grad… (#7898)

michaelroyzen committed 168 days ago

Verified be60451f

Add AGENTS.md and CLAUDE.md with project rules for AI coding agents (#7902)

delock committed 168 days ago

Verified 49a82a0d

fix: Validate fp16.loss_scale is finite and non-negative (#7889)

nathon-lee committed 169 days ago

Verified 63eeb114

Fix Evoformer's multi-arch dispatch root cause (#7881)

tohtana committed 169 days ago

Verified 784cc26e

Fix async_io ops building error on Huawei Ascend NPU (#7894)

huangyifan0610 committed 169 days ago

Verified f88d0f8d

double reduction user-friendly error (#7895)

stas00 committed 171 days ago

Verified b6346bf0

[Bloom] Fix hangs of bloom test (#7890)

k-artem committed 172 days ago

Verified d9a4aadb

Suppress see_memory_usage logs (#7891)

sfc-gh-truwase committed 174 days ago

Verified 285cae30

Fix hook count performance regression from v0.18.5 (#7886)

tohtana committed 176 days ago

Verified 6c59d544

Add document section explaining autocast nesting (#7883)

tohtana committed 177 days ago

Verified 4dba1e20

Remove amp() from abstract accelerator (#7879)

delock committed 179 days ago

Verified a41a96b1

XPU use stock pytorch instead of Intel Extension for PyTorch (#7877)

delock committed 179 days ago

Verified d8e15da4

Fix import deepspeed crash on PyTorch v2.3 + Python 3.12 (#7875)

tohtana committed 180 days ago

Verified 04d69cc6

Older