[Kernel][ROCm] Native W4A16 kernel for AMD RDNA3 (gfx1100) — fp16 + bf16 (#41394)

Signed-off-by: JartX <sagformas@epdcenter.es>
This commit is contained in:
JartX
2026-05-29 11:04:40 +00:00
committed by GitHub
parent 60a7a2214f
commit 0cff0741ff
13 changed files with 3846 additions and 0 deletions
+45
View File
@@ -650,6 +650,51 @@ def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor, bit: int) -> None
torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
def gptq_gemm_rdna3(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_qzeros: torch.Tensor,
b_scales: torch.Tensor,
b_g_idx: torch.Tensor,
use_v2_format: bool,
) -> torch.Tensor:
return torch.ops._rocm_C.gptq_gemm_rdna3(
a, b_q_weight, b_qzeros, b_scales, b_g_idx, use_v2_format
)
if hasattr(torch.ops, "_rocm_C") and hasattr(torch.ops._rocm_C, "gptq_gemm_rdna3"):
@register_fake("_rocm_C::gptq_gemm_rdna3")
def _gptq_gemm_rdna3_fake(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_qzeros: torch.Tensor,
b_scales: torch.Tensor,
b_g_idx: torch.Tensor,
use_v2_format: bool,
) -> torch.Tensor:
return torch.empty(
(a.size(0), b_q_weight.size(1)), dtype=a.dtype, device=a.device
)
if hasattr(torch.ops, "_rocm_C") and hasattr(torch.ops._rocm_C, "gptq_gemm_rdna3_wmma"):
@register_fake("_rocm_C::gptq_gemm_rdna3_wmma")
def _gptq_gemm_rdna3_wmma_fake(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_qzeros: torch.Tensor,
b_scales: torch.Tensor,
b_g_idx: torch.Tensor,
use_v2_format: bool,
) -> torch.Tensor:
return torch.empty(
(a.size(0), b_q_weight.size(1)), dtype=a.dtype, device=a.device
)
if hasattr(torch.ops._C, "allspark_w8a16_gemm"):
@register_fake("_C::allspark_w8a16_gemm")