[Kernel] Support GLM-5 dimensions for TRT-LLM ragged MLA prefill (#43525)

Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Mohammad Miadh Angkad
2026-06-16 20:49:47 +00:00
committed by GitHub
co-authored by Matthew Bonanni
parent b9684d99e9
commit 9d4dc4ca2f
10 changed files with 200 additions and 77 deletions
+10 -2
View File
@@ -765,7 +765,8 @@ def test_backend_correctness(
if not backends_to_test:
pytest.skip(f"No backends support kv_cache_dtype={kv_cache_dtype}")
# Skip prefill backends that can't satisfy capability/deps/R1 constraints.
# Skip prefill backends that can't satisfy capability/deps/dimension constraints.
from vllm.v1.attention.backends.mla.prefill.base import MLADimensions
from vllm.v1.attention.backends.mla.prefill.selector import (
MLAPrefillSelectorConfig,
)
@@ -773,7 +774,14 @@ def test_backend_correctness(
try:
prefill_invalid_reasons = prefill_backend.get_class().validate_configuration(
current_platform.get_device_capability(),
MLAPrefillSelectorConfig(dtype=torch.bfloat16, is_r1_compatible=True),
MLAPrefillSelectorConfig(
dtype=torch.bfloat16,
mla_dimensions=MLADimensions(
qk_nope_head_dim=128,
qk_rope_head_dim=64,
v_head_dim=128,
),
),
)
except ImportError:
prefill_invalid_reasons = ["ImportError"]