[dependabot]

Updates the requirements on deepspeed to permit the latest version.

Release notes

Sourced from deepspeed's releases.

v0.19.8 Patch Release

What's Changed

• Deprecate leftovers of compression library by @​hwchen2017 in deepspeedai/DeepSpeed#8535
• Update version.txt after 0.19.7 release by @​loadams in deepspeedai/DeepSpeed#8555
• Deprecate loco zero++ by @​hwchen2017 in deepspeedai/DeepSpeed#8534
• Deprecate sparse_gradients by @​pengdurice in deepspeedai/DeepSpeed#8504
• docs: stop Codex false-alarm on DCO sign-off by @​delock in deepspeedai/DeepSpeed#8561
• Raise Modal sandbox timeout to absorb full-suite runtime by @​jinyouzhi in deepspeedai/DeepSpeed#8571
• Derive the autocast test backend and scope the bf16 gate to NCCL by @​delock in deepspeedai/DeepSpeed#8559
• Skip heavy CI for docs-only PRs touching root markdown by @​delock in deepspeedai/DeepSpeed#8560
• Deprecate APEX amp by @​hwchen2017 in deepspeedai/DeepSpeed#8564
• Document sub_group_size and correct its default to 1e9 by @​shravanibnikam in deepspeedai/DeepSpeed#8567
• Deprecate legacy curriculum learning by @​hwchen2017 in deepspeedai/DeepSpeed#8566
• [muon] Per-head Newton-Schulz for attention projections by @​alanhuangyoo in deepspeedai/DeepSpeed#8384
• Upgrade yapf to v0.43.0 to fix EOFError failure in CI by @​jinyouzhi in deepspeedai/DeepSpeed#8574
• [AutoTP] Add support for vocab-parallel LM head and related configurations by @​jinyouzhi in deepspeedai/DeepSpeed#8309
• Remove Progressive Layer Dropping (PLD) support by @​hwchen2017 in deepspeedai/DeepSpeed#8562
• Support Muon in BF16_Optimizer with FP32 gradient accumulation by @​0z5a in deepspeedai/DeepSpeed#8526
• Register native pinned host memory with the CANN runtime for NPU by @​VenusTZZ in deepspeedai/DeepSpeed#8532
• implement register_host_memory and unregister_host_memory for supa by @​frozenleaves in deepspeedai/DeepSpeed#8537
• Remove elastic training by @​pengdurice in deepspeedai/DeepSpeed#8563
• Use int64 program ids in the SwiGLU and AutoEP fused restore Triton kernels by @​pengdurice in deepspeedai/DeepSpeed#8591
• Describe the GPTBigCode and Yuan shared-QK layouts by @​Achyuthan-S in deepspeedai/DeepSpeed#8575
• Fix HybridEngine OPT generation with modern cache interfaces by @​0z5a in deepspeedai/DeepSpeed#8523
• [muon] Per-head Muon for linear-attention layers: read head geometry from the owning module by @​alanhuangyoo in deepspeedai/DeepSpeed#8436
• [AutoTP] Support eager row-parallel output head training by @​0z5a in deepspeedai/DeepSpeed#8538
• Release AutoEP router-logit caches on checkpoint replay exits by @​yh0903 in deepspeedai/DeepSpeed#8581
• Fix memory leak and NULL-deref in shm_initialize by @​tritsystem in deepspeedai/DeepSpeed#8565
• Fix silent gradient-bucket drop on CPU in ZeRO-1/2 non-contiguous reduction by @​delock in deepspeedai/DeepSpeed#8382
• Add opt-in AutoEP Python GC policy by @​yh0903 in deepspeedai/DeepSpeed#8451
• [AutoTP] Add optional Liger vocabulary-parallel cross entropy by @​0z5a in deepspeedai/DeepSpeed#8540
• Overlap AutoEP split planning with token packing by @​yh0903 in deepspeedai/DeepSpeed#8359
• Normalize Gram Newton-Schulz inputs before casting to fp16 by @​vineethsaivs in deepspeedai/DeepSpeed#8533
• Remove the older TiledLoss and UlyssesSPFwdLossBwdWithLogits by @​vineethsaivs in deepspeedai/DeepSpeed#8587
• Wait for all async all-gathers in coalesced ZeRO-3 parameter gather by @​delock in deepspeedai/DeepSpeed#8539
• Skip narrow dtypes in the mixed-precision cast by @​sfc-gh-abkulkarni in deepspeedai/DeepSpeed#8416
• AutoEP: drop the redundant float32 copy in the default combine by @​pengdurice in deepspeedai/DeepSpeed#8621
• Remove stale compression test from HPU CI by @​nathon-lee in deepspeedai/DeepSpeed#8616
• Skip fp16 ZeroPP tests on accelerators without fp16 support by @​jinyouzhi in deepspeedai/DeepSpeed#8610
• AutoEP: share one DeepEP buffer across MoE layers instead of one per layer by @​yh0903 in deepspeedai/DeepSpeed#8557
• Implement graph operations for NPU accelerator by @​VenusTZZ in deepspeedai/DeepSpeed#8605
• Add opt-in regional compilation around AutoEP layers by @​yh0903 in deepspeedai/DeepSpeed#8380
• Fix pipe tests failures on CPU by @​jinyouzhi in deepspeedai/DeepSpeed#8611
• Run Muon once per optimizer step under ZeRO-3, not once per micro-batch by @​alanhuangyoo in deepspeedai/DeepSpeed#8600
• Fix 2 CPU failures in cpu-torch-latest by @​jinyouzhi in deepspeedai/DeepSpeed#8629
• Fix Muon optimizer under ZeRO CPU offload and bound gather buffers by @​jinyouzhi in deepspeedai/DeepSpeed#8464
• Loosen the late-iteration gradient tolerance in the checkpointing test by @​delock in deepspeedai/DeepSpeed#8648
• Add affine AutoEP checkpoint placement by @​jinyouzhi in deepspeedai/DeepSpeed#8544
• Fix SequenceTiledCompute backward for empty trailing shards by @​taking-lying-flat in deepspeedai/DeepSpeed#8434
• fix: Remove the unused ROCm Apex clone, build, and PYTHONPATH setup. by @​nathon-lee in deepspeedai/DeepSpeed#8617

... (truncated)

Commits

• 0eb3a8b Enable CUDA Graph for continuous rollouts (#8732)
• 9169472 Add KLXPU (KUNLUNXIN XPU) accelerator (#8799)
• cc7b512 Release process-group caches before CPU/Gloo worker shutdown (#8846)
• 0fa8494 Add deepspeed.comm.pg_utils.is_parallel for optional parallel groups (#8815)
• 2b43788 Fix TestLionConfigs on the CPU accelerator and run it on every runner (#8833)
• a4811fb [CI] Restore Modal-first torch-latest routing (#8849)
• 036230e Add vocabulary-parallel fused linear cross entropy (#8792)
• f07f19f Describe the CodeGen fused-QKV layout as a block selection (#8706)
• 53f59db Add vocabulary-parallel logprobs, entropy and KL ops (#8791)
• 3b7de87 Release GIL while destroying AIO handles (#8832)
• Additional commits viewable in compare view