microsoft/DeepSpeed

Pull Requests Commits

fix(rollout): fix CUDA graph capture producing incorrect logits

delock committed 36 days ago

918b6904

expand question list to 200+

delock committed 36 days ago

5e4e8d12

rollout: add FlashInfer kernel manager and vLLM compat shim

delock committed 37 days ago

f9a73e06

fix: AutoTP _replace_module uses full hierarchical path for pattern matching

delock committed 38 days ago

60dd200a

rollout: fix hybrid engine device index bug, vllm import path, and smoke configs

delock committed 39 days ago

db8289ad

remove weight_bridge: vLLM handles TP slicing internally

delock committed 39 days ago

d1454ccf

fix: remove student_model_path from RolloutConfig in tests

delock committed 39 days ago

b82b5a3a

test: add CPU unit tests for VLLMRollout and HybridEngineRollout

delock committed 39 days ago

94a5551f

move rollout tests from examples/opsd/tests to tests/unit/runtime/rollout

delock committed 39 days ago

088fcf96

refactor: remove naive HF generate path from HybridEngineRollout

delock committed 39 days ago

0886a10b

feat: add vllm_python config to specify vLLM interpreter

delock committed 39 days ago

d4bc9509

feat: abstract weight transfer backend (gdr/http) for VLLMRollout

delock committed 39 days ago

9a262aee

refactor: drop redundant _rlhf_ prefix from VLLMRollout private methods

delock committed 39 days ago

affcf439

clear and restore forward hooks before and after rollout

delock committed 39 days ago

065e3844

move vllm rollout into deepspeed/runtime

delock committed 40 days ago

7e14117c

vllmrollout support and move vllmrollout into deepspeed/runtime/rollout

delock committed 40 days ago

03cbb06e

add bench scripts

delock committed 41 days ago

6340d853

move trainer to deepspeed/runtime/rlhf/trainer

delock committed 41 days ago

aa015931

fix: DynamicCache compat for transformers>=5 in _cb_replace_slot

delock committed 41 days ago

9a943f2d

restore forward hooks after rollout

delock committed 43 days ago

7b352113

set capture mode to global

delock committed 43 days ago

7c8bbb62

fix: enable CUDA graph capture with DeepSpeed hybrid engine

delock committed 46 days ago

2d34fd3e

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

delock committed 46 days ago

9b4d823e

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

delock committed 46 days ago

bfac5063

Rename student_engine -> engine, sync_weights_from_student -> sync_weights

delock committed 47 days ago

e816c457

Add kv_trim_threshold parameter to HybridEngineRollout

delock committed 47 days ago

739079b5

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

delock committed 47 days ago

1abb504d

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

delock committed 47 days ago

a897889e

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

PKUWZP committed 49 days ago

6384396b

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

PKUWZP committed 49 days ago

837787a0

Older