[PR #349] [MERGED] Sync nv_dev with upstream #347 #381

Closed
opened 2026-10-26 21:17:04 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/349
Author: @Barry-Delaney
Created: 6/1/2026
Status: ✅ Merged
Merged: 6/2/2026
Merged by: @RayWang96

Base: nv_dev ← Head: user/jinshik/nv_dev_347


📝 Commits (3)

📊 Changes

29 files changed (+546 additions, -431 deletions)

View changed files

📝 csrc/apis/attention.hpp (+2 -2)
📝 csrc/apis/einsum.hpp (+1 -2)
📝 csrc/apis/gemm.hpp (+1 -2)
📝 csrc/apis/mega.hpp (+4 -2)
📝 csrc/jit_kernels/heuristics/mega_moe.hpp (+71 -35)
📝 csrc/jit_kernels/impls/runtime_utils.hpp (+3 -2)
📝 csrc/jit_kernels/impls/sm100_fp8_fp4_mega_moe.hpp (+12 -6)
📝 csrc/jit_kernels/impls/sm90_fp8_gemm_1d2d.hpp (+2 -1)
📝 csrc/jit_kernels/impls/smxx_fp8_fp4_paged_mqa_logits.hpp (+3 -2)
📝 deep_gemm/include/deep_gemm/comm/barrier.cuh (+3 -3)
📝 deep_gemm/include/deep_gemm/impls/sm100_bf16_gemm.cuh (+6 -6)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp4_mqa_logits.cuh (+8 -8)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp4_paged_mqa_logits.cuh (+9 -9)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_gemm_1d1d.cuh (+8 -10)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh (+178 -198)
📝 deep_gemm/include/deep_gemm/impls/sm100_fp8_paged_mqa_logits.cuh (+3 -4)
📝 deep_gemm/include/deep_gemm/impls/sm90_bf16_gemm.cuh (+10 -4)
📝 deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d1d.cuh (+5 -2)
📝 deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d2d.cuh (+6 -1)
📝 deep_gemm/include/deep_gemm/layout/sym_buffer.cuh (+3 -0)

...and 9 more files

📄 Description

Merges upstream 88965b0 (deepseek-ai PR #347: "Multiple updates and refactorings") into nv_dev (c
urrently ac1f285 = #316 merge via PR #328 + PR #342 IMA guard). Brings PR #347's FP8/FP4 Mega MoE r
efactor, BF16 accumulation/output expansion, paged-MQA scheduler refactor, and assorted fixes. Merge
base is 891d57b; the merge also pulls 714dd1a ("Update test_mega_moe.py").

Topology follows the established nv_dev convention (git merge --no-ff <upstream_sha>, preserving th
e upstream commit as the second parent), matching PR #314 / #328.


🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/deepseek-ai/DeepGEMM/pull/349 **Author:** [@Barry-Delaney](https://github.com/Barry-Delaney) **Created:** 6/1/2026 **Status:** ✅ Merged **Merged:** 6/2/2026 **Merged by:** [@RayWang96](https://github.com/RayWang96) **Base:** `nv_dev` ← **Head:** `user/jinshik/nv_dev_347` --- ### 📝 Commits (3) - [`714dd1a`](https://github.com/deepseek-ai/DeepGEMM/commit/714dd1a4a980f7937a74343d19a8eba4fe321480) Update test_mega_moe.py - [`88965b0`](https://github.com/deepseek-ai/DeepGEMM/commit/88965b078186ee7510ab9fc4f1d5ebc19adfa8d1) Multiple updates and refactorings (#347) - [`58af0be`](https://github.com/deepseek-ai/DeepGEMM/commit/58af0be80213ac1bd2993c6ff6f1c97f66e574f8) Merge upstream commit 88965b0 (#347: Multiple updates and refactorings) into nv_dev ### 📊 Changes **29 files changed** (+546 additions, -431 deletions) <details> <summary>View changed files</summary> 📝 `csrc/apis/attention.hpp` (+2 -2) 📝 `csrc/apis/einsum.hpp` (+1 -2) 📝 `csrc/apis/gemm.hpp` (+1 -2) 📝 `csrc/apis/mega.hpp` (+4 -2) 📝 `csrc/jit_kernels/heuristics/mega_moe.hpp` (+71 -35) 📝 `csrc/jit_kernels/impls/runtime_utils.hpp` (+3 -2) 📝 `csrc/jit_kernels/impls/sm100_fp8_fp4_mega_moe.hpp` (+12 -6) 📝 `csrc/jit_kernels/impls/sm90_fp8_gemm_1d2d.hpp` (+2 -1) 📝 `csrc/jit_kernels/impls/smxx_fp8_fp4_paged_mqa_logits.hpp` (+3 -2) 📝 `deep_gemm/include/deep_gemm/comm/barrier.cuh` (+3 -3) 📝 `deep_gemm/include/deep_gemm/impls/sm100_bf16_gemm.cuh` (+6 -6) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp4_mqa_logits.cuh` (+8 -8) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp4_paged_mqa_logits.cuh` (+9 -9) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_gemm_1d1d.cuh` (+8 -10) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh` (+178 -198) 📝 `deep_gemm/include/deep_gemm/impls/sm100_fp8_paged_mqa_logits.cuh` (+3 -4) 📝 `deep_gemm/include/deep_gemm/impls/sm90_bf16_gemm.cuh` (+10 -4) 📝 `deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d1d.cuh` (+5 -2) 📝 `deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d2d.cuh` (+6 -1) 📝 `deep_gemm/include/deep_gemm/layout/sym_buffer.cuh` (+3 -0) _...and 9 more files_ </details> ### 📄 Description Merges upstream `88965b0` (deepseek-ai PR #347: "Multiple updates and refactorings") into `nv_dev` (c urrently `ac1f285` = #316 merge via PR #328 + PR #342 IMA guard). Brings PR #347's FP8/FP4 Mega MoE r efactor, BF16 accumulation/output expansion, paged-MQA scheduler refactor, and assorted fixes. Merge base is `891d57b`; the merge also pulls `714dd1a` ("Update test_mega_moe.py"). Topology follows the established nv_dev convention (`git merge --no-ff <upstream_sha>`, preserving th e upstream commit as the second parent), matching PR #314 / #328. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
bsakaguchi 2026-10-26 21:17:04 +00:00
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepGEMM#381
No description provided.