Comfy-Org/comfy-kitchen

Fork 0

mirror of https://github.com/Comfy-Org/comfy-kitchen.git synced 2026-11-04 23:37:42 +00:00

[PR #130] Tune gfx1151 INT8 deep-K tile selection #164

New issue

Open

opened 2026-10-26 22:21:02 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:21:02 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/130
Author: @Yasei-no-otoko
Created: 8/22/2026
Status: 🔄 Open

Base: main ← Head: codex/optimize-int8-convrot-gfx1151

📝 Commits (1)

0a2b1f4 Tune gfx1151 INT8 deep-K tile selection

📊 Changes

3 files changed (+107 additions, -14 deletions)

View changed files

📝 comfy_kitchen/backends/hip/gemm_wmma.h (+44 -14)
📝 comfy_kitchen/backends/hip/ops/gemm_int8.hip (+17 -0)
📝 tests/test_hip_wmma.py (+46 -0)

📄 Description

Summary

Build directly on current main, including #122's shared occupancy-aware WMMA selector.
Add a 256x128x128 INT8 specialization only for the measured gfx1151 target when M >= 512, N >= 128, K >= 14336, the grid covers the device, and either K >= 2N or the grid divides evenly across WGPs.
Derive the gfx target and WGP count from the GEMM allocation in the shared WMMA helper, so the specialized and fallback selectors cannot inspect different process-current devices.
Keep unmeasured gfx110x/gfx115x targets and RDNA 4 on #122's selector.
Keep benchmark tooling only on the separate operational branch; this PR changes the shared WMMA device probe, the INT8 launcher, and its tests.

Why this predicate

The original K >= 2N predicate was re-tested. It preserves the H3 contraction but is too broad without a depth floor: representative smaller-K shapes regressed by 0.5% to 20.8% on gfx1151. A plain K >= 14336 gate also selected the 256-row tile on some wide-N grids where the shared selector was faster. K = 14336 is the MiniMax-H3 mlp_down contraction depth.

The final predicate retains the H3/deep-K case and adds the 256-row tile only for complete WGP rounds. Compared with the previous K >= 14336 gate (HIP events, alternating fresh processes, median of run medians):

Shape `(M,N,K)`	Previous	Final	Change
`(512,8192,14336)`	3.6192 ms	3.1618 ms	-12.6%
`(512,9216,14336)`	3.7262 ms	3.6161 ms	-3.0%
`(1024,8192,14336)`	6.5110 ms	6.3050 ms	-3.2%
H3 `(3802,5376,14336)`	same tile	15.1003 ms	neutral

The current allocation-derived shared probe recheck measured H3 mlp_down at 15.1154 ms on gfx1151, within the prior run spread. Tested HIP outputs are bit-identical between selectors.

E2E on gfx1151

MiniMax-H3, same models/prompt/seed/settings, 124 frames at 24 fps, 20 steps:

Full 1344x768 fresh-process comparison after rebasing: current main 2119 s; #130 2093 s (-1.23%).
Fast selector iteration at 416x224, H.264 High + AAC-LC:

Selector	E2E
Previous `K >= 14336` gate	145.34 s
Original `K >= 2N` candidate	144.21 s
Final WGP + `K >= 2N` gate	145.35 s

The spread is below 0.8% and within fresh-process noise; the H3 shape selects the same tile in all three. Every file decodes to exactly 124 frames at 416x224/24 fps plus 5.167 s stereo AAC, with identical decoded video and audio SHA-256 hashes. The final review change only narrows other architectures and centralizes the device probe; it does not change gfx1151 tile selection or kernel math.

Architecture policy

RDNA 3 (gfx110x): #122 selector retained; compile validated, not performance-measured locally.
RDNA 3.5: specialization enabled only on measured gfx1151; gfx1150, gfx1152, and gfx1153 retain #122.
RDNA 4 (gfx12xx): #122 selector retained; compile and contributor runtime validated for gfx1200.

Validation

32-thread HIP build for gfx1100, gfx1151, gfx1200, and gfx1201
Targeted H3/deep-K plus shared FP8 paths: 30 passed
test_hip_wmma.py, test_int8.py, and test_int8_input_act.py on gfx1151: 474 passed, 62 skipped
ruff check and git diff --check
gfx1151 H3 mlp_down HIP-event sanity check on final head: 15.1154 ms

_{🔄 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/Comfy-Org/comfy-kitchen/pull/130 **Author:** [@Yasei-no-otoko](https://github.com/Yasei-no-otoko) **Created:** 8/22/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `codex/optimize-int8-convrot-gfx1151` --- ### 📝 Commits (1) - [`0a2b1f4`](https://github.com/Comfy-Org/comfy-kitchen/commit/0a2b1f49951648f887f0b1ac597ec36ae0d93b25) Tune gfx1151 INT8 deep-K tile selection ### 📊 Changes **3 files changed** (+107 additions, -14 deletions) <details> <summary>View changed files</summary> 📝 `comfy_kitchen/backends/hip/gemm_wmma.h` (+44 -14) 📝 `comfy_kitchen/backends/hip/ops/gemm_int8.hip` (+17 -0) 📝 `tests/test_hip_wmma.py` (+46 -0) </details> ### 📄 Description ## Summary - Build directly on current `main`, including #122's shared occupancy-aware WMMA selector. - Add a `256x128x128` INT8 specialization only for the measured `gfx1151` target when `M >= 512`, `N >= 128`, `K >= 14336`, the grid covers the device, and either `K >= 2N` or the grid divides evenly across WGPs. - Derive the gfx target and WGP count from the GEMM allocation in the shared WMMA helper, so the specialized and fallback selectors cannot inspect different process-current devices. - Keep unmeasured gfx110x/gfx115x targets and RDNA 4 on #122's selector. - Keep benchmark tooling only on the separate [operational branch](https://github.com/Yasei-no-otoko/comfy-kitchen/tree/codex/ops-int8-pr130-pr127-pr129-gfx1151); this PR changes the shared WMMA device probe, the INT8 launcher, and its tests. ## Why this predicate The original `K >= 2N` predicate was re-tested. It preserves the H3 contraction but is too broad without a depth floor: representative smaller-K shapes regressed by 0.5% to 20.8% on gfx1151. A plain `K >= 14336` gate also selected the 256-row tile on some wide-N grids where the shared selector was faster. `K = 14336` is the MiniMax-H3 `mlp_down` contraction depth. The final predicate retains the H3/deep-K case and adds the 256-row tile only for complete WGP rounds. Compared with the previous `K >= 14336` gate (HIP events, alternating fresh processes, median of run medians): | Shape `(M,N,K)` | Previous | Final | Change | |---|---:|---:|---:| | `(512,8192,14336)` | 3.6192 ms | 3.1618 ms | -12.6% | | `(512,9216,14336)` | 3.7262 ms | 3.6161 ms | -3.0% | | `(1024,8192,14336)` | 6.5110 ms | 6.3050 ms | -3.2% | | H3 `(3802,5376,14336)` | same tile | 15.1003 ms | neutral | The current allocation-derived shared probe recheck measured H3 `mlp_down` at 15.1154 ms on gfx1151, within the prior run spread. Tested HIP outputs are bit-identical between selectors. ## E2E on gfx1151 MiniMax-H3, same models/prompt/seed/settings, 124 frames at 24 fps, 20 steps: - Full 1344x768 fresh-process comparison after rebasing: current `main` 2119 s; #130 2093 s (-1.23%). - Fast selector iteration at 416x224, H.264 High + AAC-LC: | Selector | E2E | |---|---:| | Previous `K >= 14336` gate | 145.34 s | | Original `K >= 2N` candidate | 144.21 s | | Final WGP + `K >= 2N` gate | 145.35 s | The spread is below 0.8% and within fresh-process noise; the H3 shape selects the same tile in all three. Every file decodes to exactly 124 frames at 416x224/24 fps plus 5.167 s stereo AAC, with identical decoded video and audio SHA-256 hashes. The final review change only narrows other architectures and centralizes the device probe; it does not change gfx1151 tile selection or kernel math. ## Architecture policy - RDNA 3 (`gfx110x`): #122 selector retained; compile validated, not performance-measured locally. - RDNA 3.5: specialization enabled only on measured `gfx1151`; `gfx1150`, `gfx1152`, and `gfx1153` retain #122. - RDNA 4 (`gfx12xx`): #122 selector retained; compile and contributor runtime validated for gfx1200. ## Validation - [x] 32-thread HIP build for gfx1100, gfx1151, gfx1200, and gfx1201 - [x] Targeted H3/deep-K plus shared FP8 paths: 30 passed - [x] `test_hip_wmma.py`, `test_int8.py`, and `test_int8_input_act.py` on gfx1151: 474 passed, 62 skipped - [x] `ruff check` and `git diff --check` - [x] gfx1151 H3 `mlp_down` HIP-event sanity check on final head: 15.1154 ms --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-26 22:21:02 +00:00

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

Comfy-Org/comfy-kitchen#164

No description provided.

Rows
Columns