Skip to content

[Triton/Gluon] [Config] Drop no-op kpack from RDNA GEMM configs - #5917

Merged
azaidy merged 4 commits into
ROCm:mainfrom
hyjuunn:rdna-drop-kpack
Sep 29, 2026
Merged

azaidy merged 4 commits into
ROCm:mainfrom
hyjuunn:rdna-drop-kpack

Conversation

@hyjuunn

@hyjuunn hyjuunn commented Sep 28, 2026 •

Copy link
Copy Markdown
Contributor

Motivation

Follow up on #5557, as suggested by @amd-xavierwang and requested by @azaidy.

kpack is a Triton AMD backend compiler option used in MFMA lowering. It does not affect the generated code for the gfx1151/gfx1201 GEMM configs in this change, but is still present in 52 config files. Some buckets contain kpack, while others in the same file do not.

This PR extends the kpack cleanup applied to gfx1100 blockscale configs in #5217 to the remaining gfx1151/gfx1201 GEMM configs.

Technical Details

Removed 397 kpack entries from 52 JSON files:

Architecture Files Removed entries
gfx1151 11 135
gfx1201 41 262

Affected families:

  • gfx1151: gemm_a8w8 (1 file), gemm_a8w8_blockscale (10).
  • gfx1201: gemm_a8w8 (5), gemm_a8w8_blockscale (32), gemm_a8w8_blockscale_preshuffled (1), batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant (3).

The diff only removes kpack entries and the trailing comma of the preceding property where needed. Also verified that:

  • Each modified JSON equals the original with only kpack removed.
  • All other values, key order, indentation and final newlines are preserved.
  • Changes are limited to the intended config files.
  • git diff --check passes.

matrix_instr_nonkdim is retained. It is listed among the required GEMM keys in aiter/ops/triton/configs/CLAUDE.md.

No kernel, Python, test or documentation files are changed.

Test Plan

Environment

  • Image: rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0
  • Digest: sha256:4449f856653602317e4101a76fce599c7fcd58ccec2e539951fce5f73083179e
  • PyTorch: 2.10.0+rocm7.2.4.git3d3aa833
  • AMD Triton: 3.8.0+amd.rocm7.2.0.git7cb7b059, installed using install_triton.sh
  • Container ROCm: 7.2.4

GPU tests and benchmarks compared upstream main at 8253efc4 with the kpack-only commit 64e67f5e, using separate Triton caches. The restored branch at ef6646fc8 has exactly the same Git tree as 64e67f5e; the results below are from that original validation. Focused fused-consumer checks were rerun after restoration (see below):

  • Radeon AI PRO R9700 — gfx1201
  • Radeon 8060S / Strix Halo — gfx1151

Validation

  1. Offline ISA comparison (no GPU). Compiled every changed config with and without kpack and compared TTIR, TTGIR, LLVM IR, and AMDGCN. As a control, kpack=1 vs kpack=2 on gfx942 produces different output.
  2. Existing pytest: test_gemm_a8w8_blockscale.py (Triton path), test_gemm_a8w8.py, and test_batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant.py, compared per nodeid between baseline and branch.
  3. Targeted harness. Covers every changed file and every M_* bucket, plus the first M past each bucket boundary. It calls the public wrappers with config=None, compares results against the torch reference and the baseline output, and verifies that the config loaded on the branch contains no kpack. This includes the K<512 shapes that the existing blockscale test skips.
  4. Performance. Ran the existing op_tests/op_benchmarks/triton benches for representative shapes, 5 runs per revision, each in a separate process.

Test Result

  • Offline ISA: identical at every stage for all 135 gfx1151 and all 312 gfx1201 config entries. The gfx942 control differs, as expected. The 447 entries include 397 entries with kpack removed and 50 controls that already omitted it. The final comparison used a fresh process per compile call.

  • Harness: all 52 changed files were exercised on their target GPU.

    GPU files cases vs torch ref (baseline / branch) baseline vs branch output
    gfx1201 41 571 571/571 / 571/571 bit-identical (max abs diff 0)
    gfx1151 11 250 250/250 / 250/250 bit-identical (max abs diff 0)

    No config loaded on the branch contains kpack. The K<512 cases (40 on gfx1201, 26 on gfx1151) pass as well.

  • pytest: per-nodeid results are identical between baseline and branch on both GPUs.

    GPU Test pass / fail / skip (baseline = branch)
    gfx1201 blockscale (Triton) 226 / 4 / 8
    gfx1201 a8w8 753 / 0 / 340
    gfx1201 batched 392 / 0 / 0
    gfx1151 blockscale (Triton) 115 / 115 / 8
    gfx1151 a8w8 753 / 0 / 340
    gfx1151 batched 0 / 392 / 0

    All failing nodeids were reproduced on the upstream main baseline (8253efc4) in the same test environment. No new failing nodeids were introduced.

    • gfx1151 blockscale: 115 shuffle=True cases fail because the preshuffled default config is absent.
    • gfx1151 batched: 373 cases fail because the default config is absent, and 19 fail with OOM during input or torch reference preparation. The failure categories match between baseline and branch.
    • gfx1201 blockscale: the four failures comprise one assertion failure and three native crashes, all in preshuffle cases with M=1:
      • N=7168, K=768: GROUP_K must equal BLOCK_SIZE_K.
      • N=8192/16384/65536, K=1536: GPU page fault and native process abort.
  • Fused-consumer checks after restoration: on gfx1201 at ef6646fc8, all 4 mul/add, 2 FP8 BMM+RoPE and 2 split-cat cases passed. These call existing test functions directly. The 6 mul/add and BMM+RoPE cases passed on main, failed with missing-key errors after the matrix-option removal, and now pass again. Loaded configs retain matrix_instr_nonkdim=16 and omit kpack.

  • Performance:

    • gfx1201: median latency changed by −0.46% to +0.87% across 9 shapes.
    • gfx1151: initial median latency changes ranged from −3.68% to +4.41% across 8 shapes. Only A8W8 64×4096×4096 exceeded +3%. A longer re-measurement of that shape (100 ms warmup, 1000 ms rep, 10 interleaved pairs) gave +0.38%. The initial slowdown was not reproduced. The kernel's executable instructions are identical; raw AMDGCN differs only in debug checkout paths and the corresponding string-offset comments.
    • Clocks were not pinned on either machine.
  • CI does not exercise the gfx1151/gfx1201 configs changed here, so validation on these targets was performed manually.

Submission Checklist

@github-actions

Copy link
Copy Markdown
Contributor

🏷️ CI Guide

Runs automatically on every PR:

  • ✅ Pre-checks (submodule verification, code formatting)
  • ✅ Aiter op tests (gfx942 + gfx950)
  • ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label Tests
ci:gfx1250-ffm-triton Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
multigpu Aiter multi-GPU tests on the 8-GPU runner
ci:sglang SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 5917 --add-label <label>

One backend per PR:
A PR changes one kernel backend: [Triton/Gluon] (Triton and Gluon count as one), [HIP], [ASM], [CK], [OPUS] or [FlyDSL]. If the title ends up with two backend tags, split the PR -- as stacked pull requests when one part cannot merge without the other.

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to stop the title rewrites; labels stay in sync either way.

@hyjuunn hyjuunn changed the title [Triton] [Config] Drop no-op kpack from RDNA GEMM configs [Triton/Gluon] [Config] Drop no-op kpack from RDNA GEMM configs Sep 28, 2026
@hyjuunn hyjuunn changed the title [Triton/Gluon] [Config] Drop no-op kpack from RDNA GEMM configs [Triton/Gluon] [Config] Drop no-op kpack and matrix_instr_nonkdim from RDNA GEMM configs Sep 28, 2026
@hyjuunn
hyjuunn marked this pull request as ready for review September 28, 2026 20:51
@hyjuunn
hyjuunn requested review from a team and a lite review from Copilot September 28, 2026 20:51
@amd-xavierwang

Copy link
Copy Markdown
Contributor

LGTM

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot review overview

🟡 Changes recommended

Fused blockscale paths require matrix_instr_nonkdim, and the configuration contract also needs resolution for its omission.

Review effort: Lite
Findings: 3 High severity

Open (3)
What changed in this PR

This PR removes kpack and matrix_instr_nonkdim entries from RDNA Triton GEMM configuration files.

Changes:

  • Cleans gfx1100, gfx1151, and gfx1201 GEMM configurations.
  • Preserves other tuning values and formatting.
  • Updates standard, blockscale, preshuffled, and batched configurations.
File Change
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8/​GEMM-A8W8-N=6144-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8/​GEMM-A8W8-N=4096-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8/​GEMM-A8W8-N=4096-K=14336.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8/​GEMM-A8W8-N=28672-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8/​DEFAULT.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=8192-K=8192.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=8192-K=32768.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=8192-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=7168-K=256.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=7168-K=2304.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=7168-K=2048.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=7168-K=18432.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=7168-K=16384.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=6144-K=2048.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=6144-K=1536.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=512-K=7168.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=512-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4608-K=7168.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=7168.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=512.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=256.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=36864-K=7168.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=3584-K=512.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=32768-K=8192.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=32768-K=512.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=3072-K=6144.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=3072-K=1536.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=2624-K=6144.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=24576-K=1536.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=2112-K=7168.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=2048-K=2048.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=16384-K=1536.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=1536-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=1024-K=8192.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=1024-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale/​DEFAULT.json Removes targeted keys; fused-wrapper compatibility remains unresolved.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale_preshuffled/​GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=6144-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale_preshuffled/​GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=4096-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale_preshuffled/​GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=4096-K=12288.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale_preshuffled/​GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=24576-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​gemm_a8w8_blockscale_preshuffled/​DEFAULT.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant/​DEFAULT.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant/​BATCHED_GEMM-A8W8-A_PER_TOKEN_GROUP_PREQUANT_W_PER_BATCHED_TENSOR_QUANT-N=512-K=128.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1201/​triton/​gemm/​batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant/​BATCHED_GEMM-A8W8-A_PER_TOKEN_GROUP_PREQUANT_W_PER_BATCHED_TENSOR_QUANT-N=128-K=512.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8/​DEFAULT.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=8192-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=512-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=256.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=2048-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=16384-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=1536-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1151/​triton/​gemm/​gemm_a8w8_blockscale/​DEFAULT.json Removes targeted keys; fused-wrapper compatibility remains unresolved.
aiter/​ops/​triton/​configs/​gfx1100/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=8192-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1100/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=512-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1100/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=256.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1100/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=4096-K=1024.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1100/​triton/​gemm/​gemm_a8w8_blockscale/​GEMM-A8W8_BLOCKSCALE-N=1536-K=4096.json Removes targeted configuration keys.
aiter/​ops/​triton/​configs/​gfx1100/​triton/​gemm/​gemm_a8w8_blockscale/​DEFAULT.json Removes targeted keys; fused-wrapper compatibility remains unresolved.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

@hyjuunn
hyjuunn marked this pull request as draft September 28, 2026 21:14
@hyjuunn hyjuunn changed the title [Triton/Gluon] [Config] Drop no-op kpack and matrix_instr_nonkdim from RDNA GEMM configs [Triton] [Config] Drop no-op kpack from RDNA GEMM configs Sep 28, 2026
@hyjuunn

hyjuunn commented Sep 28, 2026

Copy link
Copy Markdown
Contributor Author

Restored matrix_instr_nonkdim in ef6646f and kept this PR limited to kpack removal (52 JSON files, 397 entries).

The shared configs also feed fused_gemm_a8w8_blockscale_mul_add, which requires the key as a constexpr argument, and fused_fp8_bmm_rope_cat_and_cache_mla, whose wrapper indexes it directly. Removing it caused a TypeError and a KeyError, respectively. On gfx1201, the 4 mul/add and 2 BMM+RoPE cases passed on main, failed after removal, and all pass after restoration.

The final tree is identical to the previously validated kpack-only commit 64e67f5e. Keeping matrix-option removal would require changes to those consumers; those changes are outside this config-only PR. The description is updated accordingly.

@hyjuunn
hyjuunn marked this pull request as ready for review September 28, 2026 21:34
Copilot AI review requested due to automatic review settings September 28, 2026 21:34
@github-actions github-actions Bot changed the title [Triton] [Config] Drop no-op kpack from RDNA GEMM configs [Triton/Gluon] [Config] Drop no-op kpack from RDNA GEMM configs Sep 28, 2026

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot review overview

🟢 Approval recommended

Validation found no unresolved issues or behavior changes.

Review effort: Lite
Findings: None

Resolved since last review (3)

@azaidy azaidy self-assigned this Sep 28, 2026
@azaidy
azaidy self-requested a review September 28, 2026 21:40

@azaidy azaidy left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM!

@azaidy
azaidy merged commit 9ef3e41 into ROCm:main Sep 29, 2026
104 of 108 checks passed
Boss2002n added a commit that referenced this pull request Sep 29, 2026
…e configs

#5877 added these three after #5917 cleared kpack from the other RDNA GEMM
configs. The fused BMM + RoPE launcher never passes kpack, and it is a no-op
on WMMA anyway, so removing it changes nothing.
Boss2002n added a commit that referenced this pull request Sep 29, 2026
With #5917 and the previous commit, gfx1100, gfx1101, gfx1150, gfx1151,
gfx1200 and gfx1201 carry no kpack, so the three passages saying the RDNA
trees still do are updated: kpack is a no-op on any WMMA target, only gfx1250
has entries left, and dropping one needs no retune.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants