[V1] [Hybrid] Support Minimax-Text-01 in V1 (#22151)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2025-08-08 23:08:48 -07:00
committed by GitHub
parent 3157aebb63
commit 6ade99eafa
5 changed files with 234 additions and 42 deletions
+1 -1
View File
@@ -532,7 +532,7 @@ def _linear_attn_decode_kernel(
pid_d = tl.program_id(2) # dimension block index
# Load slot index for the current batch
slot_id = tl.load(slot_idx + pid_b)
slot_id = tl.load(slot_idx + pid_b).to(tl.int64)
# Skip if slot_id is -1 (padding)
if slot_id == -1:
@@ -5,6 +5,17 @@ from vllm.distributed import divide
class MambaStateShapeCalculator:
@classmethod
def linear_attention_state_shape(
cls,
num_heads: int,
tp_size: int,
head_dim: int,
) -> tuple[tuple[int, int, int], ...]:
state_shape = (num_heads // tp_size, head_dim, head_dim)
return (state_shape, )
@classmethod
def mamba1_state_shape(
cls,