[hxbai]

DeepSeek-V4 main tracker

Updated: 2026-10-20. Tracks mainline support for the DSv4 HybridModel performance recipe.

Progress: Core foundations and all three tracked functional ports—Hash MoE, packed THD/CP, and mHC PP/VPP—are merged into main. Remaining work is performance parity.

BatchedLinear blocker: Dev #5722 and main #7971 both depend on TE #3160 for the optimized path. In #7971, the pinned TE does not provide BatchedLinear: the draft uses the einsum fallback and skips real BatchedLinear tests. Resolve the TE dependency, update the pin, recheck the final API, and pass BF16/MXFP8 GPU validation before marking the optimization ready. #7971 remains a draft.

Finished — merged into main

Completed features and supporting infrastructure

PR title and summary	Dev PR(s)	Main PR(s)
DSv4 native/fused attention and HybridModel integration	#4458, #4894, #6349, #5762	#6400, #6402, #6404, #6405
Native/fused mHC and HybridModel/MTP integration	#2943, #3828, #4624, #6172	#4531, #6401
CSA parameter precision and FP8-context isolation	#5308	#6400, #5929
Selective-FP32 runtime and optimizer lifecycle	#6124	#5929, #4547
Clamped SwiGLU across expert paths	#4481, #5130	#5940
Safe MLA RoPE and packed-position handling	#5018, #5526	#5944
Variable-length data, packing, and CP-layout foundations	#2924	#5902, #5903, #5908, #6625, #6684, #6734, #6742, #6776, #6912
THD variable-token dispatcher support	#2924, #5048	#5008
Muon FP8 parameter gather and buffer reuse	#5470	#6754
TE graph parameter-gradient clone removal	— (main-first)	#5893
NCCL EP baseline	— (main-first)	#5129
NCCL EP zero-copy, overflow handling, and MXFP8 dispatch	— (main-first)	#5735, #6229, #6524, #7195
Device-initiated GroupedLinear	#6847	#6000
Local CUDA graph memory-pool reuse	— (main-first)	#5451
Local CUDA graph stream-aliasing fix	— (main-first)	#6643
DSv4 packed THD/CP — Static context parallelism and MTP support.	#5011, #5087, #5691, #5706, #5809	#7469
Hash MoE — Routing and training integration.	#4481, #5130, #4596, #6704	#6403
mHC PP/VPP — Pipeline support for multistream residuals.	#2943, #4518, #5471	#7336

Existing main PRs — performance work

PR title and summary	Dev PR(s)	Main PR(s)
MoK backend — Dispatch, expert execution, combine, and shared-output gate; single-node GB200 functional validation passed.	#6572, #7289, #7391	#8210 (ready for review)
BatchedLinear output projection — Optimized path blocked on TE #3160.	#5722 (open)	#7971 (draft)
DSv4 FLOPs accounting — Correct reported metrics; no execution speedup.	#5142, #5358, #5762	#7972 (ready for review)
THD TE attention graphs — Scheduler-packed capture/replay; replaces #5783.	#4359, #5724	#8361 (ready for review)
mHC TE wrappers and MTP discovery — Base graph support; recompute split remains below.	#4190, #7014, #6583	#7337 (draft)
Remove dummy attention masks — Reduce graph memory.	#5131	#5173 (open)
Muon compact layout + FP8 sync — Reduce optimizer layout memory.	#5388, #5470	#5479 (open)
Output inverse RoPE reconstruction — Reduce saved activation memory.	#7036	#8399 (draft)
DP-balanced all-gather rerouting — Reduce transport-buffer memory.	#6378	#7047 (open; includes this item)

MoK dependency installation #7696 is merged. #8210 is the tracked backend port onto current main.

Remaining performance ports — dev merged

PR title and summary	Dev PR(s)	Main PR(s)
Compact BF16/MXFP8 indexer — Faster CSA scoring and Top-K.	#5992	—
Fused compressor pooling — Faster pooling forward/backward.	#5984, #7064	—
THD/CP preprocessing kernels — Fused layout and sparse-loss preprocessing.	#7064	— (#7469 provides functional THD/CP support)
mHC recompute + attention graph split — Preserve recompute memory savings with graphs.	#5841, #6661, #7385	— (#7337 covers base graphs only)
MTP halo prefetch and counter batching — Reduce MTP communication overhead.	#6246	— (#7469 provides basic rolls)
CP compressor projection before compaction — Avoid retaining compact hidden-size inputs.	#7388	—
Fixed-shape packed PP, including VPP — Skip repeated shape exchange.	#7000, #7483	—

The CC 10.7 mapping in dev #7779 is merged. Carry it with the CuTe layout-kernel port; current main does not contain the module it patches.

— = no active main PR identified that covers the remaining optimization. These are feature gaps, not a required one-PR-per-row split. Existing PRs and partial coverage are listed above; detailed code comparisons and validation evidence are kept locally.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.