[Kernel] Porting fuse_minimax_qk_norm to manual fusion (#43410)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-05-26 13:16:03 -07:00
committed by GitHub
parent 49b4882779
commit 6e503868ca
12 changed files with 262 additions and 490 deletions
@@ -39,7 +39,6 @@ from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.mamba.linear_attn import (
MiniMaxText01LinearAttention,
MiniMaxText01LinearKernel,
MiniMaxText01RMSNormTP,
clear_linear_attention_cache_for_new_sequences,
linear_attention_decode,
linear_attention_prefill_and_mix,
@@ -49,6 +48,7 @@ from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
)
from vllm.model_executor.layers.minimax_rms_norm import MiniMaxText01RMSNormTP
from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
+4 -3
View File
@@ -50,7 +50,7 @@ from vllm.model_executor.layers.linear import (
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.linear_attn import MiniMaxText01RMSNormTP
from vllm.model_executor.layers.minimax_rms_norm import MiniMaxText01RMSNormTP
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -243,8 +243,9 @@ class MiniMaxM2Attention(nn.Module):
hidden_states: torch.Tensor,
) -> torch.Tensor:
qkv, _ = self.qkv_proj(hidden_states)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
q, k = MiniMaxText01RMSNormTP.forward_qk(self.q_norm, self.k_norm, q, k)
q, k, v = MiniMaxText01RMSNormTP.forward_qkv(
self.q_norm, self.k_norm, qkv, self.q_size, self.kv_size
)
q, k = self.rotary_emb(positions, q, k)
attn_output = self.attn(q, k, v)
output, _ = self.o_proj(attn_output)