deepseek-ai/DeepGEMM

Fork 0

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

[PR #468] [CLOSED] [DeepGEMM] retire actlize v0.5.0 and use v1.0.0 only #471

New issue

Closed

opened 2026-10-26 21:18:15 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:18:15 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/468
Author: @chenchu-zs
Created: 10/12/2026
Status: ❌ Closed

Base: main ← Head: feature/chenchu.zs-dev

📝 Commits (10+)

a741d37 Deepgemm for PPU
203be8f [DeepGemm]: fix mqa_logits kv stride
728e997 [DeepGemm]: Optimize fp4 TP GEMM2 performance
1900ec2 [mqa_logits]: fix nan in int8 with bf16 output
116bfad [fused moe]: compile-time topk for moe_align, 7-item fused config
f9255c3 Fix einsum compilation error
7130433 [FP4] Fine-tune dynamic tile logic to optimize Qwen3.8 gemm1/gemm2 EP performance
4e519b5 docs: add PPU overview to README
cb4e42f [mqa_logits]: multiple updates and refactorings
6bdbd2f feat(densegemm): adaptive tile select for bf16/int8, split dense gemm from moe gemm

📊 Changes

236 files changed (+141499 additions, -3158 deletions)

View changed files

➕ .github/workflows/ci-deepgemm.yml (+49 -0)
➕ .github/workflows/deepgemm-ut-test-ppu.yml (+156 -0)
➕ .github/workflows/scripts/casecmd_810e.list (+14 -0)
➕ .github/workflows/scripts/casecmd_890p.list (+16 -0)
➕ .github/workflows/scripts/display-test-summary.py (+180 -0)
➕ .github/workflows/scripts/generate-env-info.py (+66 -0)
➕ .github/workflows/scripts/pr-deepgemm-ut.py (+212 -0)
📝 .gitignore (+11 -2)
📝 .gitmodules (+8 -3)
➕ CHANGELOG.md (+44 -0)
➖ CMakeLists.txt (+0 -44)
📝 LICENSE (+20 -0)
📝 README.md (+147 -98)
➕ csrc/apis/attention.hpp (+641 -0)
➕ csrc/apis/einsum.hpp (+175 -0)
➕ csrc/apis/gemm.hpp (+1727 -0)
➕ csrc/apis/hyperconnection.hpp (+64 -0)
➕ csrc/apis/layout.hpp (+69 -0)
➕ csrc/apis/runtime.hpp (+71 -0)
➕ csrc/indexing/main.cu (+9 -0)

...and 80 more files

📄 Description

Summary

This PR removes the legacy actlize_v0.5.0 implementation and makes
actlize_v1.0.0 the only PPU implementation in DeepGEMM-for-sail.

It removes approximately 2,900 lines of legacy code and configuration while
preserving the existing v1.0.0 behavior on PPU1.0 and PPU1.5.

Motivation

All migration-scope actlize_v1.0.0 paths have passed correctness and
performance qualification on ZW810E:

BF16 Dense and Grouped GEMM
INT8 Dense and Grouped GEMM
INT8 Einsum / BatchGemm

Maintaining v0.5.0 requires duplicated kernels, compiler include routing,
packaging logic, and an environment-variable selector.

The old implementation also cannot support functionality such as the plain
[B, M, N] Einsum output layout.

Changes

Remove actlize v0.5.0

Remove the third-party/actlize_v0.5.0 submodule.
Remove its setup, package-data, symlink, and include-path configuration.
Remove the legacy BF16 and INT8 implementation headers and runtimes.
Remove the obsolete predicated-tile iterator parameter helper.

Remove version selection

Remove DG_USE_ACTLIZE_V100 and use_actlize_v100.
Remove ActlizeLib and per-kernel actlize include selection.
Make the Python and C++ JIT compilers use actlize_v1.0.0 by default.

Preserve device-specific v1.0.0 routing

PPU1.0 uses the general v1.0.0 BF16 and INT8 runtime.
PPU1.5 keeps the specialized Dense v1.0.0 runtime.
Grouped GEMM and Einsum use v1.0.0 directly.
Standalone GEMV, FP8, DynamicTile, and other independent paths are
preserved.

Preserve recent main-branch behavior

The branch was rebased onto main@356e9a7 and preserves:

the latest stream handling
INT8 Einsum bhr,hdr->hbd output-layout support
recent MQA and paged-MQA changes

Behavior change

DG_USE_ACTLIZE_V100 is no longer supported or required.

actlize_v1.0.0 is now selected unconditionally. There is no v0.5.0
fallback.

No public Python API is changed.

Validation

Static checks

Confirmed that tracked source files no longer reference:

actlize_v0.5.0
DG_USE_ACTLIZE_V100
use_actlize_v100
ActlizeLib
the removed BF16 and INT8 legacy runtimes

Additional checks:

Python source compilation passed.
Grouped JIT format unit test passed.
Patch formatting checks passed.

ZW810E correctness

Validated from commit bd07b6b after rebuilding, with
DG_USE_ACTLIZE_V100 unset:

Workload	Result
BF16 Dense + Grouped	136/136
INT8 Dense + Grouped	386/386
INT8 Einsum	10/10

No test reported an execution error, traceback, assertion failure, numerical
mismatch, segmentation fault, or invalid test dispatch.

Performance qualification

The v1.0.0 implementation was qualified against v0.5.0 before removing the
legacy implementation:

Workload	Cases	Median delta	Worst delta
BF16 Dense + Grouped	136	-0.97%	+2.92%
INT8 Dense + Grouped	376 profiled	-5.19%	+1.28%
INT8 Einsum `bhr,hdr->bhd`	5	-3.33%	-1.58%

The cleanup does not change the selected v1.0.0 kernels. Performance was not
remeasured after deleting the unused v0.5.0 implementation.

Reviewer notes

Please focus on:

PPU1.0 routing to the general v1.0.0 runtime.
PPU1.5 routing to the specialized Dense v1.0.0 runtime.
Removal of all v0.5.0 build and packaging dependencies.
Preservation of the Einsum plain-output layout.
Preservation of standalone GEMV, FP8, and DynamicTile paths.

Out of scope

The pending actlize epilogue R2S-copy optimization is not included in this
PR and should be reviewed separately in the actlize repository.

_{🔄 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/468 **Author:** [@chenchu-zs](https://github.com/chenchu-zs) **Created:** 10/12/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `feature/chenchu.zs-dev` --- ### 📝 Commits (10+) - [`a741d37`](https://github.com/deepseek-ai/DeepGEMM/commit/a741d37ab7cd4189425b56ffc2d020bbafd52f0e) Deepgemm for PPU - [`203be8f`](https://github.com/deepseek-ai/DeepGEMM/commit/203be8f2127ce19e28bdc4a4b37cdd54f385c60a) [DeepGemm]: fix mqa_logits kv stride - [`728e997`](https://github.com/deepseek-ai/DeepGEMM/commit/728e997c2d0daca11876baf4c64a58e3b44f2138) [DeepGemm]: Optimize fp4 TP GEMM2 performance - [`1900ec2`](https://github.com/deepseek-ai/DeepGEMM/commit/1900ec29bdb0cc1f7b7baddb76f210931b71b4a3) [mqa_logits]: fix nan in int8 with bf16 output - [`116bfad`](https://github.com/deepseek-ai/DeepGEMM/commit/116bfad622b224aa227f8e4a15714fab83c93e1b) [fused moe]: compile-time topk for moe_align, 7-item fused config - [`f9255c3`](https://github.com/deepseek-ai/DeepGEMM/commit/f9255c3e44efd809bb334b0691dbc0c6a4dc439e) Fix einsum compilation error - [`7130433`](https://github.com/deepseek-ai/DeepGEMM/commit/7130433fee76e94bd8a7ccdffbbb57b674a29715) [FP4] Fine-tune dynamic tile logic to optimize Qwen3.8 gemm1/gemm2 EP performance - [`4e519b5`](https://github.com/deepseek-ai/DeepGEMM/commit/4e519b541b30feff160aed68baccd25e26165afa) docs: add PPU overview to README - [`cb4e42f`](https://github.com/deepseek-ai/DeepGEMM/commit/cb4e42f4125be15d9ff1186aecef0f9bdd1e0523) [mqa_logits]: multiple updates and refactorings - [`6bdbd2f`](https://github.com/deepseek-ai/DeepGEMM/commit/6bdbd2fb7d3391d07f0eb05ed869eb054b7de7b9) feat(densegemm): adaptive tile select for bf16/int8, split dense gemm from moe gemm ### 📊 Changes **236 files changed** (+141499 additions, -3158 deletions) <details> <summary>View changed files</summary> ➕ `.github/workflows/ci-deepgemm.yml` (+49 -0) ➕ `.github/workflows/deepgemm-ut-test-ppu.yml` (+156 -0) ➕ `.github/workflows/scripts/casecmd_810e.list` (+14 -0) ➕ `.github/workflows/scripts/casecmd_890p.list` (+16 -0) ➕ `.github/workflows/scripts/display-test-summary.py` (+180 -0) ➕ `.github/workflows/scripts/generate-env-info.py` (+66 -0) ➕ `.github/workflows/scripts/pr-deepgemm-ut.py` (+212 -0) 📝 `.gitignore` (+11 -2) 📝 `.gitmodules` (+8 -3) ➕ `CHANGELOG.md` (+44 -0) ➖ `CMakeLists.txt` (+0 -44) 📝 `LICENSE` (+20 -0) 📝 `README.md` (+147 -98) ➕ `csrc/apis/attention.hpp` (+641 -0) ➕ `csrc/apis/einsum.hpp` (+175 -0) ➕ `csrc/apis/gemm.hpp` (+1727 -0) ➕ `csrc/apis/hyperconnection.hpp` (+64 -0) ➕ `csrc/apis/layout.hpp` (+69 -0) ➕ `csrc/apis/runtime.hpp` (+71 -0) ➕ `csrc/indexing/main.cu` (+9 -0) _...and 80 more files_ </details> ### 📄 Description ## Summary This PR removes the legacy `actlize_v0.5.0` implementation and makes `actlize_v1.0.0` the only PPU implementation in DeepGEMM-for-sail. It removes approximately 2,900 lines of legacy code and configuration while preserving the existing v1.0.0 behavior on PPU1.0 and PPU1.5. ## Motivation All migration-scope `actlize_v1.0.0` paths have passed correctness and performance qualification on ZW810E: - BF16 Dense and Grouped GEMM - INT8 Dense and Grouped GEMM - INT8 Einsum / BatchGemm Maintaining v0.5.0 requires duplicated kernels, compiler include routing, packaging logic, and an environment-variable selector. The old implementation also cannot support functionality such as the plain `[B, M, N]` Einsum output layout. ## Changes ### Remove actlize v0.5.0 - Remove the `third-party/actlize_v0.5.0` submodule. - Remove its setup, package-data, symlink, and include-path configuration. - Remove the legacy BF16 and INT8 implementation headers and runtimes. - Remove the obsolete predicated-tile iterator parameter helper. ### Remove version selection - Remove `DG_USE_ACTLIZE_V100` and `use_actlize_v100`. - Remove `ActlizeLib` and per-kernel actlize include selection. - Make the Python and C++ JIT compilers use `actlize_v1.0.0` by default. ### Preserve device-specific v1.0.0 routing - PPU1.0 uses the general v1.0.0 BF16 and INT8 runtime. - PPU1.5 keeps the specialized Dense v1.0.0 runtime. - Grouped GEMM and Einsum use v1.0.0 directly. - Standalone GEMV, FP8, DynamicTile, and other independent paths are preserved. ### Preserve recent main-branch behavior The branch was rebased onto `main@356e9a7` and preserves: - the latest stream handling - INT8 Einsum `bhr,hdr->hbd` output-layout support - recent MQA and paged-MQA changes ## Behavior change `DG_USE_ACTLIZE_V100` is no longer supported or required. `actlize_v1.0.0` is now selected unconditionally. There is no v0.5.0 fallback. No public Python API is changed. ## Validation ### Static checks Confirmed that tracked source files no longer reference: - `actlize_v0.5.0` - `DG_USE_ACTLIZE_V100` - `use_actlize_v100` - `ActlizeLib` - the removed BF16 and INT8 legacy runtimes Additional checks: - Python source compilation passed. - Grouped JIT format unit test passed. - Patch formatting checks passed. ### ZW810E correctness Validated from commit `bd07b6b` after rebuilding, with `DG_USE_ACTLIZE_V100` unset: | Workload | Result | |---|---:| | BF16 Dense + Grouped | 136/136 | | INT8 Dense + Grouped | 386/386 | | INT8 Einsum | 10/10 | No test reported an execution error, traceback, assertion failure, numerical mismatch, segmentation fault, or invalid test dispatch. ### Performance qualification The v1.0.0 implementation was qualified against v0.5.0 before removing the legacy implementation: | Workload | Cases | Median delta | Worst delta | |---|---:|---:|---:| | BF16 Dense + Grouped | 136 | -0.97% | +2.92% | | INT8 Dense + Grouped | 376 profiled | -5.19% | +1.28% | | INT8 Einsum `bhr,hdr->bhd` | 5 | -3.33% | -1.58% | The cleanup does not change the selected v1.0.0 kernels. Performance was not remeasured after deleting the unused v0.5.0 implementation. ## Reviewer notes Please focus on: 1. PPU1.0 routing to the general v1.0.0 runtime. 2. PPU1.5 routing to the specialized Dense v1.0.0 runtime. 3. Removal of all v0.5.0 build and packaging dependencies. 4. Preservation of the Einsum plain-output layout. 5. Preservation of standalone GEMV, FP8, and DynamicTile paths. ## Out of scope The pending actlize epilogue R2S-copy optimization is not included in this PR and should be reviewed separately in the actlize repository. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:18:15 +00:00