v0.19.7 Patch Release

What's Changed

• Add configurable dtype for ZeRO checkpoint export by @gaoxiaomo in #8318
• Raise the per-tensor norms to norm_type when combining them by @vineethsaivs in #8313
• Fix the seq-first Ulysses all2all output layout by @vineethsaivs in #8317
• Count each module object once when aggregating flops profiler totals by @ebarkhordar in #8320
• Fallback for unsupported Hybrid Engine policies by @LiRunGuo in #8265
• [tiled mlp] reshape instead of view by @stas00 in #8348
• Add pin_empty helper for empty scratch destinations by @sfc-gh-truwase in #8332
• Fix Triton NFS detection crash when df wraps long device names by @atirna in #8259
• [AutoTP] Replace tp_shard process-wide globals with per-model AutoTPMeta by @delock in #8241
• feat(cpu-adam): add ARM SVE update kernel by @xylian86 in #8365
• Add macOS (MPS) CI workflow and a torch floor check for the MPS accelerator by @PKUWZP in #8335
• Add an opt-in DeepEP transport for the AutoEP expert all-to-all by @yh0903 in #8213
• remove the dead non-Triton attention path from TritonSelfAttention by @Anai-Guo in #8349
• perf(rollout): profile prefill and decode forwards by @nathon-lee in #8350
• DeepCompile: stabilize ZeRO-3 parameter guards by @tohtana in #8328
• fix: Only bind device id when needed, Fixes #8248 by @pengdurice in #8269
• Make the WarmupCosineLR ratio flags reach the config by @vineethsaivs in #8268
• Forward free_data through partition() instead of hardcoding True by @vineethsaivs in #8305
• Recognize Qwen3.5's RMSNorm variants in AutoTP module loading by @promptsmith1990 in #8306
• Forward barrier device_ids to communication backends by @tohtana in #8312
• Fix flops profiler counts for transposed convolutions by @vineethsaivs in #8323
• Broadcast elementwise flops from the trailing dimension by @vineethsaivs in #8324
• Replace cuda_graph assertion with explicit ValueError; fix custom_op type annotations by @chakshu-dhannawat in #8336
• Keep the elasticity batch overrides out of the caller's config dict by @alanhuangyoo in #8329
• Stop the curriculum schedule starting below min_difficulty by @vineethsaivs in #8334
• Add an opt-in fused weighted restore for AutoEP by @yh0903 in #8326
• Add nccl_version to the source-checkout torch_info fallback by @alanhuangyoo in #8383
• docs: add test discipline rules to agent guidelines by @delock in #8372
• Preallocate the static KV cache with config.head_dim by @ebarkhordar in #8389
• Skip fp16-config tests on accelerators without fp16 support by @delock in #8398
• Do not pin DDP device_ids for CPU reference models by @delock in #8399
• Use device names, not rank ids, for device placement in test helpers by @delock in #8397
• [Ulysses] Carry the KV head count per DistributedAttention by @alanhuangyoo in #8316
• [Workflow] Raise modal CI timeouts to absorb slower sandbox provisioning by @delock in #8403
• Preserve AutoEP score correction bias buffers by @poorpaper in #8369
• op_builder: use C++20 for nvcc on CUDA 13+ by @bheu in #8422
• Reshape instead of view in TiledFusedLogitsLoss by @alanhuangyoo in #8362
• Fix AutoTP + deep compile collectives silently drop when AC is on by @pengdurice in #8355
• fix(lr_schedules): make --lr_range_test_staircase an opt-in flag by @Anai-Guo in #8337
• Split the modal CI budget into acquisition and test phases by @delock in #8404
• [Workflow] run modal GPU workflows only from the merge queue by @delock in #8412
• Read rope_theta from rope_parameters in the Llama injection policy by @alanhuangyoo in #8341
• Stop the debug name maps from pinning the model they snapshot by @alanhuangyoo in #8356
• Muon runs no Newton-Schulz at ZeRO stage 0, the default: run it by @alanhuangyoo in #8442
• Filter --include against the real slots, not against itself by @vineethsaivs in #8239
• Add an opt-in ZeRO-1/2 gradient norm fast path by @yh0903 in #8331
• [AutoEP]Fix optimizer and replaced MOE parameter mismatch by @pengdurice in #8377
• Muon is silently disabled under ZeRO-3 when the model is built with zero.Init by @alanhuangyoo in #8438
• Avoid collective token preparation for AutoEP DeepEP by @yh0903 in #8423
• [MPS] Update C++ Standard in CPUAdamBuilder by @jinyouzhi in #8466
• Split DeepCompile ZeRO-3 memory scheduler by @tohtana in #8233
• Describe universal checkpoint shards as affine maps by @Achyuthan-S in #8385
• Fix comms logger KeyError when log_name is omitted by @jinyouzhi in #8267
• Fix universal checkpoint resume across AutoTP sizes by @0z5a in #8474
• feat(rollout): add continuous batching generation prototype by @nathon-lee in #8368
• Fix ZeRO parameter alignment for grouped_mm by @fwerkor in #8277
• Carry the affine scale on the replicated map, not the split by @Achyuthan-S in #8477
• [muon] Reconcile the momentum dtype when a checkpoint is restored by @alanhuangyoo in #8433
• Deprecate unused DeepSpeed features by @sfc-gh-truwase in #8490
• [muon] Keep the momentum out of steps the loss scaler discards by @alanhuangyoo in #8435
• Remove triton compatibility check for fp_quantizer by @hwchen2017 in #8492
• Muon silently discards the param groups it is given by @alanhuangyoo in #8440
• [Phase 2] Add NEON SIMD path for CPU Adam on AArch64 by @PKUWZP in #8453
• Fix sequence overlap backward gradient permutation by @taking-lying-flat in #8342
• Validate positive inference and HybridEngine max output tokens by @taking-lying-flat in #8343
• Offload a saved view when it is the last value holding its storage by @pengdurice in #8388
• Honor --include/--exclude in the SLURM launcher by @vineethsaivs in #8304
• Make wait() idempotent on AllGatherHandle and NoGatherHandle by @pengdurice in #8487
• Deprecate sparse attention by @hwchen2017 in #8493
• feat(rollout): add continuous batching profiling by @nathon-lee in #8494
• Probe the device module for train_cifar's fork_rng device entries by @delock in #8407
• Gate the offload-state memory deltas on allocator-backed stats by @delock in #8409
• Read rope_theta from rope_parameters across Inference V2 by @alanhuangyoo in #8345
• Partition AutoEP expert parameters per layer under ZeRO-3 by @tohtana in #8424
• Give Muon's momentum the dtype of the gradient it is combined with by @alanhuangyoo in #8483
• Update version.txt after 0.19.6 release by @loadams in #8333
• Add DCO sign-off to release commits by @loadams in #8527
• Emit affine maps from AutoTP layers by @Achyuthan-S in #8519
• Cache the Modal sandbox install chain in image layers by @delock in #8536
• [bugfix] _DimZeroAllToAll silently fails to send grads under torch.compile by @stas00 in #8491

New Contributors

• @atirna made their first contribution in #8259
• @Anai-Guo made their first contribution in #8349
• @promptsmith1990 made their first contribution in #8306
• @chakshu-dhannawat made their first contribution in #8336
• @poorpaper made their first contribution in #8369
• @bheu made their first contribution in #8422
• @0z5a made their first contribution in #8474
• @taking-lying-flat made their first contribution in #8342

Full Changelog: v0.19.6...v0.19.7