deepseek-ai/DeepGEMM

Fork 0

mirror of https://github.com/deepseek-ai/DeepGEMM.git synced 2026-11-05 00:47:26 +00:00

[PR #470] [CLOSED] Sync/upstream dev `e1f418c2` #477

New issue

Closed

opened 2026-10-26 21:18:19 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:18:19 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/470
Author: @deepak-kumar-neu
Created: 10/13/2026
Status: ❌ Closed

Base: nv_dev ← Head: sync/upstream-dev-e1f418c2

📝 Commits (10+)

e8147a8 Merge upstream/nv_dev at 8b1392b into midstream/nv_dev
9e4a022 Merge pull request #10 from Gregory-Pereira/sync/upstream-nv_dev-8b1392b
66081d4 Public Release 26/09 (#432)
20567a3 Port SM120 (consumer Blackwell) kernels from nv_dev (#4)
78b6900 Fix Mega MoE task info slot release ordering (#441)
650bf51 Add configurable SwiGLU alpha and beta (#5)
30b0968 Fix Mega MoE task info slot release ordering
db64865 Merge branch 'deepseek-ai:main' into dev
e49b29b Add SiTU activation to FP8/FP4 Mega MoE (#6)
9a86ae2 SM90 paged MQA logits: support kv_block=32 and next_n=4 (#7)

📊 Changes

161 files changed (+18258 additions, -7117 deletions)

View changed files

📝 .gitmodules (+3 -3)
➕ AI/2026-09-12-sm120-port-design.md (+400 -0)
➕ AI/2026-09-12-sm120-port-plan.md (+1616 -0)
📝 CMakeLists.txt (+4 -2)
📝 README.md (+37 -24)
📝 csrc/apis/attention.hpp (+290 -123)
📝 csrc/apis/config.hpp (+17 -21)
📝 csrc/apis/einsum.hpp (+116 -90)
📝 csrc/apis/gemm.hpp (+278 -193)
📝 csrc/apis/hyperconnection.hpp (+2 -8)
📝 csrc/apis/layout.hpp (+12 -21)
➕ csrc/apis/mega_gate.hpp (+234 -0)
➕ csrc/apis/mega_mhc.hpp (+287 -0)
📝 csrc/apis/mega_moe.hpp (+55 -46)
➕ csrc/apis/sm120_dispatch.hpp (+193 -0)
📝 csrc/indexing/main.cu (+1 -2)
➖ csrc/jit/cache.hpp (+0 -31)
➖ csrc/jit/compiler.hpp (+0 -369)
➖ csrc/jit/device_runtime.hpp (+0 -153)
➖ csrc/jit/handle.hpp (+0 -222)

...and 80 more files

📄 Description

Summary

Merging 13 upstream commits up till e1f418c2a4f20818221f6b0e578b4c2f634d4c3f into midstream nv_dev.

This is the DeepGEMM pin referenced by DEEPGEMM_GIT_REF in cmake/external_projects/deepgemm.cmake of nm-vllm-ent sync-v0.31.0 (vLLM 0.31.0, RHAI 3.6 GA).

Upstream source change

vLLM moved the DeepGEMM pin from deepseek-ai/DeepGEMM nv_dev to the vLLM fork (vllm-project/DeepGEMM) dev branch in vllm-project/vllm#56876. The fork diverged from deepseek-ai at 559d79f (Public release 26/07) and re-implements/ports the deepseek line (e.g. 20567a3 "Port SM120 kernels from nv_dev (#4)").

This sync therefore merges from vllm-project/DeepGEMM dev, not deepseek-ai/nv_dev.

Commits being merged (oldest → newest)

66081d4 Public Release 26/09 (#432)
20567a3 Port SM120 (consumer Blackwell) kernels from nv_dev (#4)
78b6900 Fix Mega MoE task info slot release ordering (#441)
650bf51 Add configurable SwiGLU alpha and beta (#5)
30b0968 Fix Mega MoE task info slot release ordering
db64865 Merge branch 'deepseek-ai:main' into dev
e49b29b Add SiTU activation to FP8/FP4 Mega MoE (#6)
9a86ae2 SM90 paged MQA logits: support kv_block=32 and next_n=4 (#7)
ad1f172 Restore zero-context-length guards in the SM90 paged MQA scheduler (#8)
f9d0e23 fix(sm120): correct ported GEMM and MQA execution paths (#9)
40d109b Remove Mega-Gate finite-score device trap (#11)
a6bbb80 Increase device barrier watchdog timeout to 300s (#13)
e1f418c Include cuda_fp8.h / cuda_bf16.h where their types are used (#12)

Conflict resolution

The two lines diverged at 559d79f, so the merge conflicts in 47 files. Resolution strategy: take the upstream dev tree wholesale — the merge commit records both parents (nv_dev history is preserved and reachable), but the resulting tree is taken verbatim from the pin.

Verification

git diff e1f418c2 HEAD --stat → empty (tree is byte-identical to the pin)
git log -1 --format="%p" → 9e4a022 e1f418c (both parents recorded)
Diffstat vs nv_dev: 161 files, +18,258 / −7,117

Post-merge

Tag v2.8.0+rhaiv.0 will be cut at the merge commit, following the v2.6.0 convention (#10 → 9e4a022 → v2.6.0+rhaiv.0).

_{🔄 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/470 **Author:** [@deepak-kumar-neu](https://github.com/deepak-kumar-neu) **Created:** 10/13/2026 **Status:** ❌ Closed **Base:** `nv_dev` ← **Head:** `sync/upstream-dev-e1f418c2` --- ### 📝 Commits (10+) - [`e8147a8`](https://github.com/deepseek-ai/DeepGEMM/commit/e8147a8d4e8c8b76506779d71c24af2e55426a5d) Merge upstream/nv_dev at 8b1392b into midstream/nv_dev - [`9e4a022`](https://github.com/deepseek-ai/DeepGEMM/commit/9e4a02257c9140f1635157bd92d0f0b2ea1eff55) Merge pull request #10 from Gregory-Pereira/sync/upstream-nv_dev-8b1392b - [`66081d4`](https://github.com/deepseek-ai/DeepGEMM/commit/66081d4c9c7d7c44f13fea402e5b622aa0f409c2) Public Release 26/09 (#432) - [`20567a3`](https://github.com/deepseek-ai/DeepGEMM/commit/20567a3684ed1cb0bfb44e94b0917c0e0eef3ca4) Port SM120 (consumer Blackwell) kernels from nv_dev (#4) - [`78b6900`](https://github.com/deepseek-ai/DeepGEMM/commit/78b69000794d0937b47ae3387eff7663410264d1) Fix Mega MoE task info slot release ordering (#441) - [`650bf51`](https://github.com/deepseek-ai/DeepGEMM/commit/650bf51932e4014365f0092c67de6407441070b9) Add configurable SwiGLU alpha and beta (#5) - [`30b0968`](https://github.com/deepseek-ai/DeepGEMM/commit/30b096804ba4c6aeaeb0a5a7c638d20281e4f369) Fix Mega MoE task info slot release ordering - [`db64865`](https://github.com/deepseek-ai/DeepGEMM/commit/db648650ab92c7387d6d5947ea9f2f4533ca6e6a) Merge branch 'deepseek-ai:main' into dev - [`e49b29b`](https://github.com/deepseek-ai/DeepGEMM/commit/e49b29bb2b95f9ad40495374f7041e980d69004d) Add SiTU activation to FP8/FP4 Mega MoE (#6) - [`9a86ae2`](https://github.com/deepseek-ai/DeepGEMM/commit/9a86ae2b78991f1c8e6e95945a591e1ca3b19ce7) SM90 paged MQA logits: support kv_block=32 and next_n=4 (#7) ### 📊 Changes **161 files changed** (+18258 additions, -7117 deletions) <details> <summary>View changed files</summary> 📝 `.gitmodules` (+3 -3) ➕ `AI/2026-09-12-sm120-port-design.md` (+400 -0) ➕ `AI/2026-09-12-sm120-port-plan.md` (+1616 -0) 📝 `CMakeLists.txt` (+4 -2) 📝 `README.md` (+37 -24) 📝 `csrc/apis/attention.hpp` (+290 -123) 📝 `csrc/apis/config.hpp` (+17 -21) 📝 `csrc/apis/einsum.hpp` (+116 -90) 📝 `csrc/apis/gemm.hpp` (+278 -193) 📝 `csrc/apis/hyperconnection.hpp` (+2 -8) 📝 `csrc/apis/layout.hpp` (+12 -21) ➕ `csrc/apis/mega_gate.hpp` (+234 -0) ➕ `csrc/apis/mega_mhc.hpp` (+287 -0) 📝 `csrc/apis/mega_moe.hpp` (+55 -46) ➕ `csrc/apis/sm120_dispatch.hpp` (+193 -0) 📝 `csrc/indexing/main.cu` (+1 -2) ➖ `csrc/jit/cache.hpp` (+0 -31) ➖ `csrc/jit/compiler.hpp` (+0 -369) ➖ `csrc/jit/device_runtime.hpp` (+0 -153) ➖ `csrc/jit/handle.hpp` (+0 -222) _...and 80 more files_ </details> ### 📄 Description ## Summary Merging 13 upstream commits up till `e1f418c2a4f20818221f6b0e578b4c2f634d4c3f` into midstream `nv_dev`. This is the DeepGEMM pin referenced by `DEEPGEMM_GIT_REF` in `cmake/external_projects/deepgemm.cmake` of nm-vllm-ent `sync-v0.31.0` (vLLM 0.31.0, RHAI 3.6 GA). ## Upstream source change vLLM moved the DeepGEMM pin from `deepseek-ai/DeepGEMM` `nv_dev` to the **vLLM fork** (`vllm-project/DeepGEMM`) `dev` branch in vllm-project/vllm#56876. The fork diverged from deepseek-ai at `559d79f` (Public release 26/07) and re-implements/ports the deepseek line (e.g. `20567a3` "Port SM120 kernels from nv_dev (#4)"). This sync therefore merges from `vllm-project/DeepGEMM` `dev`, not `deepseek-ai/nv_dev`. ## Commits being merged (oldest → newest) - `66081d4` Public Release 26/09 (#432) - `20567a3` Port SM120 (consumer Blackwell) kernels from nv_dev (#4) - `78b6900` Fix Mega MoE task info slot release ordering (#441) - `650bf51` Add configurable SwiGLU alpha and beta (#5) - `30b0968` Fix Mega MoE task info slot release ordering - `db64865` Merge branch 'deepseek-ai:main' into dev - `e49b29b` Add SiTU activation to FP8/FP4 Mega MoE (#6) - `9a86ae2` SM90 paged MQA logits: support kv_block=32 and next_n=4 (#7) - `ad1f172` Restore zero-context-length guards in the SM90 paged MQA scheduler (#8) - `f9d0e23` fix(sm120): correct ported GEMM and MQA execution paths (#9) - `40d109b` Remove Mega-Gate finite-score device trap (#11) - `a6bbb80` Increase device barrier watchdog timeout to 300s (#13) - `e1f418c` Include cuda_fp8.h / cuda_bf16.h where their types are used (#12) ## Conflict resolution The two lines diverged at `559d79f`, so the merge conflicts in 47 files. Resolution strategy: **take the upstream dev tree wholesale** — the merge commit records both parents (nv_dev history is preserved and reachable), but the resulting tree is taken verbatim from the pin. ## Verification - `git diff e1f418c2 HEAD --stat` → **empty** (tree is byte-identical to the pin) - `git log -1 --format="%p"` → `9e4a022 e1f418c` (both parents recorded) - Diffstat vs nv_dev: 161 files, +18,258 / −7,117 ## Post-merge Tag `v2.8.0+rhaiv.0` will be cut at the merge commit, following the v2.6.0 convention (#10 → `9e4a022` → `v2.6.0+rhaiv.0`). --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:18:19 +00:00

closed this issue
added the
pull-request
label

bsakaguchi referenced this issue

2026-10-26 21:18:21 +00:00

[PR #477] [Fix][Einsum] Reject c in SM90 fp8_einsum("bhr,hdr->bhd") to prevent silent data corruption in sm90_fp8_bmm #480

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#477

No description provided.

Rows
Columns

[PR #470] [CLOSED] Sync/upstream dev e1f418c2 #477