Skip to content

[Triton/Gluon] gemm A16w16 tune for disable amdgpu-trackers #5137

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
yanxuer-999 wants to merge 3 commits into
base: main
Choose a base branch
from
+120 −0
Open

[Triton/Gluon] gemm A16w16 tune for disable amdgpu-trackers #5137

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a32b0e5
perf(pa_decode): set waves_per_eu=2 on v1 w_dot kernels to restore ~+…
yanxuer-999 Aug 30, 2026
08fca63
Add gfx950 GEMM-A16W16 N=1280 K=8192 config: lower waves_per_eu (M_LE…
yanxuer-999 Aug 31, 2026
0832ea4
Merge branch 'main' into triton_3_8_rel_tune
yanxuer-999 Sep 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions aiter/ops/triton/attention/pa_decode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,6 +249,12 @@ def paged_attn_decode_v1(
QUERY_GRP_SZ_POW2=query_grp_sz_pow2,
KV_BLK_SZ=kv_blk_sz,
KV_BLK_SZ_POW2=kv_blk_sz,
# triton 3.8 inflated this GQA dot kernel to 202 VGPR (occ 3->2).
# Raising the VGPR budget to 256 (waves_per_eu=2) lets regalloc pipeline
# LDS reads deeper: s_waitcnt lgkmcnt(0) 72->50, s_nop 36->9, ~+42% on
# llama3 decode. NOTE: v2_w_dot is intentionally NOT changed -- there
# waves_per_eu=2 spills and regresses ~-15% on llama3-405B.
waves_per_eu=2,
)


Expand Down Expand Up @@ -536,6 +542,10 @@ def paged_attn_decode_v1_per_token_quant(
QUERY_GRP_SZ_POW2=query_grp_sz_pow2,
KV_BLK_SZ=kv_blk_sz,
KV_BLK_SZ_POW2=kv_blk_sz,
# Same VGPR-budget fix as the non-quant v1_w_dot path above: this is the
# identical GQA dot compute kernel (per-token FP8 scale loads only), so
# the occ 3->2 inflation and the waves_per_eu=2 recovery apply equally.
waves_per_eu=2,
)


Expand Down
110 changes: 110 additions & 0 deletions aiter/ops/triton/configs/gfx950/triton/gemm/gemm_a16w16/GEMM-A16W16-N=1280-K=8192.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,110 @@
{
"M_LEQ_8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 8,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 6,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"num_warps": 8,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"num_warps": 8,
"num_stages": 3,
"waves_per_eu": 1,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
},
"M_LEQ_2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": null,
"NUM_KSPLIT": 1
},
"any": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"num_warps": 8,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": null,
"NUM_KSPLIT": 1
}
}
Loading