[Attention] Enable masked MHA for GLM-5 head dimensions (#53785)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Matthew Bonanni
2026-08-27 21:17:12 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 3a9bfc209f
commit de9250ac9e
10 changed files with 527 additions and 74 deletions
+32 -6
View File
@@ -1143,12 +1143,26 @@ def main():
]
else:
variants = available_variants
model_sweep = getattr(args, "model_parameter_sweep", None)
if model_sweep is None:
num_q_heads_values = [args.num_q_heads]
elif model_sweep.param_name == "num_q_heads":
num_q_heads_values = model_sweep.values
else:
raise ValueError(
"mha_vs_mqa mode only supports a num_q_heads model parameter sweep"
)
benchmark_specs = [
(spec, num_q_heads)
for spec in args.batch_specs
for num_q_heads in num_q_heads_values
]
masked_mha_max_seq_len = getattr(
args, "sparse_mla_masked_mha_max_seq_len", None
)
formatter = ResultsFormatter(console)
total = 0
for spec in args.batch_specs:
for spec, _ in benchmark_specs:
q_len = max(request.q_len for request in parse_batch_spec(spec))
for variant_label, _, _ in variants:
if (
@@ -1164,7 +1178,7 @@ def main():
total += len(backends)
with tqdm(total=total, desc="Benchmarking") as pbar:
for spec in args.batch_specs:
for spec, num_q_heads in benchmark_specs:
q_len = max(request.q_len for request in parse_batch_spec(spec))
for backend in backends:
for variant_label, force_mqa, mha_mode in variants:
@@ -1178,12 +1192,17 @@ def main():
and q_len > masked_mha_max_seq_len
):
continue
result_backend = f"{backend}_{variant_label}"
if model_sweep is not None:
result_backend = model_sweep.get_label(
result_backend, num_q_heads
)
config = BenchmarkConfig(
backend=f"{backend}_{variant_label}",
backend=result_backend,
batch_spec=spec,
num_layers=args.num_layers,
head_dim=args.head_dim,
num_q_heads=args.num_q_heads,
num_q_heads=num_q_heads,
num_kv_heads=args.num_kv_heads,
block_size=args.block_size,
device=args.device,
@@ -1211,7 +1230,7 @@ def main():
# run_mla_benchmark needs the real backend name
from mla_runner import run_mla_benchmark as run_mla
run_label = f"{backend}_{variant_label} {spec}"
run_label = f"{backend}_{variant_label} {num_q_heads}h {spec}"
pbar.set_postfix_str(run_label)
try:
@@ -1248,7 +1267,14 @@ def main():
# Display results with variant labels as separate "backends"
console.print("\n[bold green]MHA vs MQA Results:[/]")
variant_backends = [f"{b}_{v}" for b in backends for v, _, _ in variants]
variant_backends = []
for num_q_heads in num_q_heads_values:
for backend in backends:
for variant, _, _ in variants:
label = f"{backend}_{variant}"
if model_sweep is not None:
label = model_sweep.get_label(label, num_q_heads)
variant_backends.append(label)
formatter.print_table(all_results, variant_backends)
# Handle model parameter sweep mode
@@ -0,0 +1,59 @@
mode: mha_vs_mqa
model:
name: "glm-5.x"
num_layers: 78
num_q_heads: 64
num_kv_heads: 1
head_dim: 576
kv_lora_rank: 512
qk_nope_head_dim: 192
qk_rope_head_dim: 64
v_head_dim: 256
block_size: 128
max_model_len: 131072
model_parameter_sweep:
param_name: "num_q_heads"
values: [64, 32, 16, 8]
label_format: "{backend}_{value}h"
batch_specs:
- "q256s2048"
- "q512s2048"
- "q1024s2048"
- "q1536s2048"
- "q2048"
- "q256s4096"
- "q512s4096"
- "q1024s4096"
- "q1536s4096"
- "q2048s4096"
- "q4096"
- "q256s8192"
- "q512s8192"
- "q1024s8192"
- "q1536s8192"
- "q2048s8192"
- "q4096s8192"
- "q8192"
- "q256s16384"
- "q512s16384"
- "q1024s16384"
- "q1536s16384"
- "q2048s16384"
- "q4096s16384"
- "q8192s16384"
- "q16384"
- "q256s32768"
- "q512s32768"
- "q1024s32768"
- "q1536s32768"
- "q2048s32768"
- "q4096s32768"
- "q8192s32768"
- "q16384s32768"
- "q32768"
backends: ["FLASHMLA_SPARSE", "FLASHINFER_MLA_SPARSE"]
device: "cuda:0"
profile_memory: false
sparse_mla_masked_mha_max_seq_len: 131072
sparse_mla_mha_variants: ["masked_mha", "mqa"]
sparse_mla_topk_pattern: "random"
@@ -313,6 +313,9 @@ _BACKEND_PROPERTIES = {
"FLASHMLA_SPARSE": {
"query_format": "concat", # Single concatenated tensor (vs tuple)
},
"FLASHINFER_MLA_SPARSE": {
"query_format": "concat", # Single concatenated tensor (vs tuple)
},
}
@@ -1184,6 +1187,35 @@ def _run_mla_benchmark_batched(
)
# Run each benchmark with the shared impl
for config, threshold, num_splits in configs_with_params:
force_masked_mha = (
is_sparse and getattr(config, "sparse_mla_mha_mode", "auto") == "masked"
)
if force_masked_mha:
impl.masked_mha_available = True
from vllm.model_executor.layers.attention.sparse_mla_attention import (
GLOBAL_TOPK_MASK_MAX_BYTES,
_topk_mask_shape,
)
requests = parse_batch_spec(config.batch_spec)
max_query_len = max(request.q_len for request in requests)
mask_shape = _topk_mask_shape(
len(requests), max_query_len, max_query_len
)
mask_words = max(
GLOBAL_TOPK_MASK_MAX_BYTES // torch.int32.itemsize,
mask_shape[0] * mask_shape[1] * mask_shape[2],
)
if (
builder_instance.topk_mask_workspace is None
or builder_instance.topk_mask_workspace.numel() != mask_words
):
builder_instance.topk_mask_workspace = torch.zeros(
mask_words,
dtype=torch.int32,
device=device,
)
# Set threshold for this benchmark (FlashAttn/FlashMLA only)
original_threshold = None
effective_threshold = threshold