[ROCm][FP8][Kernel] FP8 quantization fused into Custom Paged Attention (#17139)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
This commit is contained in:
+2
-1
@@ -117,13 +117,14 @@ def paged_attention_rocm(
|
||||
kv_cache_dtype: str,
|
||||
k_scale: torch.Tensor,
|
||||
v_scale: torch.Tensor,
|
||||
fp8_out_scale: Optional[torch.Tensor] = None,
|
||||
) -> None:
|
||||
torch.ops._rocm_C.paged_attention(out, exp_sum, max_logits, tmp_out, query,
|
||||
key_cache, value_cache, num_kv_heads,
|
||||
scale, block_tables, seq_lens,
|
||||
query_start_loc, block_size, max_seq_len,
|
||||
alibi_slopes, kv_cache_dtype, k_scale,
|
||||
v_scale)
|
||||
v_scale, fp8_out_scale)
|
||||
|
||||
|
||||
def mla_decode_kvcache_cpu(
|
||||
|
||||
Reference in New Issue
Block a user