Skip to content

feat(gemm_a8w8_blockscale_bpreshuffle): cold-cache retune triton routing for DeepSeek MLA decode-M #4264

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

Merged
valarLip merged 1 commit into from
Jul 16, 2026
Merged

feat(gemm_a8w8_blockscale_bpreshuffle): cold-cache retune triton routing for DeepSeek MLA decode-M #4264

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2,810 changes: 1,405 additions & 1,405 deletions aiter/configs/model_configs/a8w8_blockscale_bpreshuffle_tuned_gemm_dsv3.csv
View file
Open in desktop

Large diffs are not rendered by default.

1,360 changes: 680 additions & 680 deletions aiter/configs/model_configs/dsv4_a8w8_blockscale_bpreshuffle_tuned_gemm.csv
View file
Open in desktop

Large diffs are not rendered by default.

36 changes: 18 additions & 18 deletions aiter/ops/triton/configs/gemm/gfx950-GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=2048-K=7168.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,76 +12,76 @@
"NUM_KSPLIT": 4
},
"M_LEQ_4": {
"BLOCK_SIZE_M": 8,
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
"NUM_KSPLIT": 8
},
"M_LEQ_8": {
"BLOCK_SIZE_M": 4,
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"waves_per_eu": 2,
"num_stages": 3,
"waves_per_eu": 1,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
"NUM_KSPLIT": 8
},
"M_LEQ_16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
"NUM_KSPLIT": 8
},
"M_LEQ_32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
"NUM_KSPLIT": 8
},
"M_LEQ_64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
"NUM_KSPLIT": 8
},
"M_LEQ_128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
"NUM_KSPLIT": 8
},
"M_LEQ_256": {
"BLOCK_SIZE_M": 64,
Expand All @@ -107,4 +107,4 @@
"cache_modifier": null,
"NUM_KSPLIT": 1
}
}
}
22 changes: 11 additions & 11 deletions aiter/ops/triton/configs/gemm/gfx950-GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=7168-K=16384.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"M_LEQ_8": {
"BLOCK_SIZE_M": 8,
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
Expand All @@ -13,12 +13,12 @@
},
"M_LEQ_16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 3,
"waves_per_eu": 8,
"num_warps": 2,
"num_stages": 2,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 8
Expand All @@ -29,20 +29,20 @@
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 3,
"waves_per_eu": 4,
"num_stages": 2,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 8
},
"M_LEQ_64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 1,
"num_stages": 2,
"waves_per_eu": 1,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 8
Expand Down
98 changes: 98 additions & 0 deletions aiter/ops/triton/configs/gemm/gfx950-GEMM-A8W8_BLOCKSCALE_PRESHUFFLED-N=7168-K=4096.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,98 @@
{
"M_LEQ_4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 8
},
"M_LEQ_8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 8
},
"M_LEQ_16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 8
},
"M_LEQ_32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 3,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
},
"M_LEQ_64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 4
},
"M_LEQ_128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 2
},
"M_LEQ_256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 2,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 2
},
"any": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 4,
"matrix_instr_nonkdim": 16,
"cache_modifier": ".cg",
"NUM_KSPLIT": 1
}
}
60 changes: 31 additions & 29 deletions op_tests/test_gemm_a8w8_blockscale.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,26 @@ def run_gemm_bpreshuffle(x, weightshuffle, x_scale, w_scale, dtype=dtypes.bf16):
)


@perftest(num_iters=TEST_NUM_ITERS)
def run_triton(x, weightshuffle, x_scale, w_scale, dtype=dtypes.bf16, backend=None):
# Direct call into the triton preshuffle kernel, mirroring the dispatch in
# gemm_a8w8_blockscale_bpreshuffle: reshape the (n, k) preshuffled weight to
# (n // 16, k * 16) and pass the transposed x_scale.
from aiter.ops.triton.gemm.basic.gemm_a8w8_blockscale import (
gemm_a8w8_blockscale_preshuffle,
)

n, k = weightshuffle.shape
return gemm_a8w8_blockscale_preshuffle(
x,
weightshuffle.reshape(n // 16, k * 16),
x_scale,
w_scale,
dtype=dtype,
backend=backend,
)


@benchmark()
def test_gemm(dtype, m, n, k, ck_preshuffle=True):
ret = {}
Expand Down Expand Up @@ -97,6 +117,17 @@ def test_gemm(dtype, m, n, k, ck_preshuffle=True):
ret[f"{tag} err"] = err_asm
ret["asm/ck"] = avg_c / avg_b

# Triton path requires a preshuffled weight. When not preshuffled we simply omit
# these columns; pd.DataFrame NaN-fills them for those rows in the summary.
if ck_preshuffle:
d, avg_d = run_triton(x, gemm_weight, x_scale_t, w_scale, dtype)
err_triton = checkAllclose(a, d, msg="triton", catastrophic_check=True)
ret["triton us"] = avg_d
ret["triton TFLOPS"] = m * n * k * 2 / avg_d / 1e6
ret["triton TB/s"] = (x.nbytes + weight.nbytes) / avg_d / 1e6
ret["triton err"] = err_triton
ret["triton/ck"] = avg_d / avg_b

return ret


Expand Down Expand Up @@ -201,35 +232,6 @@ def test_splitk_correctness(m=4, n=2112, k=7168, dtype=dtypes.bf16, splitK=1):
"-m",
type=int,
nargs="*",
choices=[
1,
2,
4,
8,
16,
32,
64,
96,
128,
160,
192,
224,
256,
288,
320,
352,
384,
416,
448,
480,
512,
1024,
2048,
4096,
6144,
8192,
10240,
],
default=[
1,
2,
Expand Down
Loading