[ROCm][Perf] Fused MoE W4A16 HIP kernel for AMD RDNA3 (gfx1100) (#44075)

Signed-off-by: JartX <sagformas@epdcenter.es>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
JartX
2026-06-06 15:30:39 -05:00
committed by GitHub
co-authored by TJian
parent fa27d4e9cf
commit 062b05ff3a
11 changed files with 1906 additions and 5 deletions
+53
View File
@@ -695,6 +695,59 @@ if hasattr(torch.ops, "_rocm_C") and hasattr(torch.ops._rocm_C, "gptq_gemm_rdna3
)
def moe_gptq_gemm_rdna3(
a: torch.Tensor,
c: torch.Tensor,
b_q_weight: torch.Tensor,
b_scales: torch.Tensor,
b_qzeros: torch.Tensor,
topk_weights: torch.Tensor,
sorted_token_ids: torch.Tensor,
expert_ids: torch.Tensor,
num_tokens_post_padded: torch.Tensor,
top_k: int,
block_size_m: int,
mul_topk_weight: bool,
output_topk: int = 0,
) -> None:
torch.ops._rocm_C.moe_gptq_gemm_rdna3(
a,
c,
b_q_weight,
b_scales,
b_qzeros,
topk_weights,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
top_k,
block_size_m,
mul_topk_weight,
output_topk,
)
if hasattr(torch.ops, "_rocm_C") and hasattr(torch.ops._rocm_C, "moe_gptq_gemm_rdna3"):
@register_fake("_rocm_C::moe_gptq_gemm_rdna3")
def _moe_gptq_gemm_rdna3_fake(
a: torch.Tensor,
c: torch.Tensor,
b_q_weight: torch.Tensor,
b_scales: torch.Tensor,
b_qzeros: torch.Tensor,
topk_weights: torch.Tensor,
sorted_token_ids: torch.Tensor,
expert_ids: torch.Tensor,
num_tokens_post_padded: torch.Tensor,
top_k: int,
block_size_m: int,
mul_topk_weight: bool,
output_topk: int = 0,
) -> None:
return
if hasattr(torch.ops._C, "allspark_w8a16_gemm"):
@register_fake("_C::allspark_w8a16_gemm")