[AMD][kimik3][ROCm][Perf] Fuse MLA q/kv RMSNorm in AMD Kimi-K3 MLA wrapper (#52494)

Signed-off-by: Rita Brugarolas Brufau <rita.brugarolasbrufau@amd.com>
This commit is contained in:
Rita Brugarolas
2026-09-04 10:14:51 +00:00
committed by GitHub
parent fd4a151262
commit 3ff4f02dfe
2 changed files with 123 additions and 2 deletions
+3 -2
View File
@@ -36,7 +36,7 @@ from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateDtypeCalculator,
MambaStateShapeCalculator,
)
from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper
from vllm.model_executor.layers.mla import MLAModules
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead,
@@ -63,6 +63,7 @@ from vllm.model_executor.models.utils import (
)
from vllm.models.kimi_k3.amd.kda import KimiK3DeltaAttention
from vllm.models.kimi_k3.amd.latent_moe_runner import ROCmLatentMoERunner
from vllm.models.kimi_k3.amd.mla import KimiK3MultiHeadLatentAttentionWrapper
from vllm.models.kimi_k3.amd.ops.attn_res import attn_res
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig
@@ -438,7 +439,7 @@ class KimiMLAAttention(nn.Module):
topk_indices_buffer=None,
g_proj=getattr(self, "g_proj", None),
)
self.mla_attn = MultiHeadLatentAttentionWrapper(
self.mla_attn = KimiK3MultiHeadLatentAttentionWrapper(
self.hidden_size,
self.num_local_heads,
self.scaling,