[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
+9
View File
@@ -18,6 +18,15 @@ void wvSplitKQ(const at::Tensor& in_a, const at::Tensor& in_b,
const at::Tensor& scale_a, const at::Tensor& scale_b,
const int64_t CuCount);
torch::Tensor gptq_gemm_rdna3(torch::Tensor a, torch::Tensor b_q_weight,
torch::Tensor b_qzeros, torch::Tensor b_scales,
torch::Tensor b_g_idx, bool use_v2_format);
torch::Tensor gptq_gemm_rdna3_wmma(torch::Tensor a, torch::Tensor b_q_weight,
torch::Tensor b_qzeros,
torch::Tensor b_scales,
torch::Tensor b_g_idx, bool use_v2_format);
void paged_attention(
torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,