deepspeedai/DeepSpeed

Pull Requests Commits

Merge branch 'master' into mps-neon

PKUWZP committed 3 days ago

Verified 48db5400

Muon runs no Newton-Schulz at ZeRO stage 0, the default: run it (#8442)

alanhuangyoo committed 4 days ago

Verified cbd303eb

Let torch pick the C++ standard for the MPS CPU Adam build

PKUWZP committed 4 days ago

1e0da01b

Stop the debug name maps from pinning the model they snapshot (#8356)

alanhuangyoo committed 5 days ago

Verified e9680c7d

Read rope_theta from rope_parameters in the Llama injection policy (#8341)

alanhuangyoo committed 5 days ago

Verified 6474bc55

Add NEON SIMD path for CPU Adam on AArch64

PKUWZP committed 5 days ago

dfd95919

[Workflow] run modal GPU workflows only from the merge queue (#8412)

delock committed 5 days ago

Verified 1e90aa9f

Split the modal CI budget into acquisition and test phases (#8404)

delock committed 6 days ago

Verified 666720b4

fix(lr_schedules): make --lr_range_test_staircase an opt-in flag (#8337)

Anai-Guo committed 6 days ago

Verified 8d4074cb

Fix AutoTP + deep compile collectives silently drop when AC is on (#8355)

pengdurice committed 6 days ago

Verified 92269f43

Reshape instead of view in TiledFusedLogitsLoss (#8362)

alanhuangyoo committed 6 days ago

Verified 71199367

op_builder: use C++20 for nvcc on CUDA 13+ (#8422)

bheu committed 6 days ago

Verified abf3dfa4

Preserve AutoEP score correction bias buffers (#8369)

poorpaper committed 8 days ago

Verified 32110a95

[Workflow] Raise modal CI timeouts to absorb slower sandbox provisioning (#8403)

delock committed 9 days ago

Verified 05daf059

[Ulysses] Carry the KV head count per DistributedAttention (#8316)

alanhuangyoo committed 9 days ago

Verified d4ed1f14

Use device names, not rank ids, for device placement in test helpers (#8397)

delock committed 10 days ago

Verified 4fd3c825

Do not pin DDP device_ids for CPU reference models (#8399)

delock committed 10 days ago

Verified d3a1b685

Skip fp16-config tests on accelerators without fp16 support (#8398)

delock committed 10 days ago

Verified 493dafa0

Preallocate the static KV cache with config.head_dim (#8389)

ebarkhordar committed 10 days ago

Verified e74c7071

docs: add test discipline rules to agent guidelines (#8372)

delock committed 10 days ago

Verified f15c990d

Add nccl_version to the source-checkout torch_info fallback (#8383)

alanhuangyoo committed 11 days ago

Verified 56de5705

Add an opt-in fused weighted restore for AutoEP (#8326)

yh0903 committed 11 days ago

Verified 80f19f30

Stop the curriculum schedule starting below min_difficulty (#8334)

vineethsaivs committed 11 days ago

Verified 8e09ed2a

Keep the elasticity batch overrides out of the caller's config dict (#8329)

alanhuangyoo committed 11 days ago

Verified c7eed159

Replace cuda_graph assertion with explicit ValueError; fix custom_op type annotations (#8336)

chakshu-dhannawat committed 11 days ago

Verified 26b2d995

Broadcast elementwise flops from the trailing dimension (#8324)

vineethsaivs committed 11 days ago

Verified b8bfd260

Fix flops profiler counts for transposed convolutions (#8323)

vineethsaivs committed 11 days ago

Verified 7fabe473

Forward barrier device_ids to communication backends (#8312)

tohtana committed 11 days ago

Verified 177c5aae

Forward free_data through partition() instead of hardcoding True (#8305)

vineethsaivs committed 11 days ago

Verified 9666fe5f

Recognize Qwen3.5's RMSNorm variants in AutoTP module loading (#8306)

promptsmith1990 committed 11 days ago

Verified 8e64a09b

Older