microsoft/DeepSpeed

Pull Requests Commits

perf(tp): reduce-scatter + all-gather in inference RowParallel

delock committed 11 days ago

deafce7b

feat(tp): duck-typed attention projection check for AutoTP layers

delock committed 11 days ago

25f4bb88

feat(tp): accelerator can_graph_capture_comm + TP-aware static cache

delock committed 11 days ago

3c439b52

refactor: abstract is_stream_capturing into the accelerator interface

delock committed 16 days ago

6647fb30

fix: replace tensor.is_cuda with a cuda-build guard in the cache update

delock committed 16 days ago

7e284fb5

style: clang-format, ignore-cuda markers for the format checks

delock committed 16 days ago

28587d3d

Merge branch 'master' into gma/qwen_ki_expr

delock committed 16 days ago

Verified f45c69c3

Add a nightly multi-rank CPU suite (#8806)

delock committed 16 days ago

Verified 3bd1b671

Build CPU optimizer ops without CUDA and fix the CPU Lion SIMD update (#8813)

hwchen2017 committed 16 days ago

Verified ca913e43

[CI] Run torch-latest only on AWS (#8825)

tohtana committed 16 days ago

Verified ca9cdd9e

fix(kernels): fused_add_norm row indexing — rows 1+ were never written

delock committed 17 days ago

10ccba42

Reduce chunked cross-entropy memory test footprint (#8824)

tohtana committed 17 days ago

Verified 5f266e26

docs+log: scope semantics and fallback visibility (review follow-ups)

delock committed 17 days ago

cbdc7272

test(segment_ki): direct parity vs the HF formulas, outputs and gradients

delock committed 17 days ago

3f80354e

fix(kernels): stable softplus in gdn_gates for strongly negative x

delock committed 17 days ago

f762a164

fix(rollout): pass explicit position tensors on the graph path

delock committed 17 days ago

d65cf114

revert: batched GEMV kernels measured slower than cuBLAS at b=8

delock committed 17 days ago

2925b719

feat(kernels): fused GDN z-gated output norm

delock committed 17 days ago

593e986f

feat(kernels): batched dual_gemv_silu_mul for b>1 MLP rows

delock committed 17 days ago

375ca1b0

feat(kernels): batched quad_gemv writes the concatenated GDN layout

delock committed 17 days ago

85721c49

feat(kernels): fuse q-side norm, RoPE, and output gate into decode_attn

delock committed 17 days ago

0825a691

[bug fix] ZeRO2: preserve fp32 grad accumulation precision in partition storage (#8749)

stas00 committed 18 days ago

Verified 3b809329

feat(kernels): fused add+norm at any batch size

delock committed 18 days ago

dfd03c63

feat(kernels): batched decode step for greedy b>1

delock committed 19 days ago

c067d7bc

feat(kernels): batched decode attention

delock committed 19 days ago

1e6951e4

Use C++20 for XPU op builders on torch 2.12+ (#8803)

jinyouzhi committed 19 days ago

Verified dcd8595c

AutoEP: do not slice DeepEP buffers that already hold exactly the arrived rows (#8747)

yh0903 committed 19 days ago

Verified b2047562

Add a pr-checklist agent skill and an AI contribution policy (#8744)

PKUWZP committed 20 days ago

Verified 1c862f7d

Add an opt-in fused gate+up grouped GEMM for AutoEP experts (#8733)

yh0903 committed 20 days ago

Verified 3a656d3e

Replay AutoEP's DeepEP dispatch layout in non-reentrant checkpoint recompute (#8731)

yh0903 committed 20 days ago

Verified 9ec13539

Newer Older