Load tuned fused_moe_lora shrink and expand kernel configs separately (#27435)

Signed-off-by: Yu Gong <yu3.gong@gmail.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
yugong333
2025-11-04 18:27:35 +08:00
committed by GitHub
co-authored by Jee Jee Li
parent 4022a9d279
commit 2ec401bc39
9 changed files with 911 additions and 125 deletions
+2 -1
View File
@@ -479,7 +479,8 @@ class PunicaWrapperBase(PunicaWrapperABC):
num_tokens_post_padded: torch.Tensor,
max_lora_rank: int,
top_k_num: int,
config,
shrink_config,
expand_config,
adapter_enabled: torch.Tensor,
mul_routed_weight=False,
):
+16 -6
View File
@@ -367,7 +367,8 @@ class PunicaWrapperGPU(PunicaWrapperBase):
num_tokens_post_padded: torch.Tensor,
max_lora_rank: int,
top_k_num: int,
config,
shrink_config,
expand_config,
adapter_enabled: torch.Tensor,
mul_routed_weight=False,
):
@@ -388,10 +389,19 @@ class PunicaWrapperGPU(PunicaWrapperBase):
top_k_num,
lora_ids,
adapter_enabled,
config["BLOCK_SIZE_M"],
config["BLOCK_SIZE_N"],
config["BLOCK_SIZE_K"],
config["GROUP_SIZE_M"],
config.get("SPLIT_K", 1),
shrink_config.get("BLOCK_SIZE_M", 64),
shrink_config.get("BLOCK_SIZE_N", 64),
shrink_config.get("BLOCK_SIZE_K", 32),
shrink_config.get("GROUP_SIZE_M", 8),
shrink_config.get("NUM_WARPS", 4),
shrink_config.get("NUM_STAGES", 3),
shrink_config.get("SPLIT_K", 1),
expand_config.get("BLOCK_SIZE_M", 64),
expand_config.get("BLOCK_SIZE_N", 64),
expand_config.get("BLOCK_SIZE_K", 32),
expand_config.get("GROUP_SIZE_M", 8),
expand_config.get("NUM_WARPS", 4),
expand_config.get("NUM_STAGES", 3),
expand_config.get("SPLIT_K", 1),
mul_routed_weight,
)