[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:
co-authored by
Matthew Bonanni
parent
b9684d99e9
commit
9d4dc4ca2f
@@ -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"]
|
||||
|
||||
Reference in New Issue
Block a user