[nv] add pcp support in dsv3.2 (#52046)

Signed-off-by: Summer Yang <girasoleyang@gmail.com>
This commit is contained in:
Summer Yang
2026-08-18 14:59:08 -07:00
committed by GitHub
parent d3fafe0c27
commit 5f7a20b316
6 changed files with 376 additions and 119 deletions
@@ -299,7 +299,7 @@ def sparse_attn_indexer(
kv_cache: torch.Tensor,
q_quant: torch.Tensor,
q_scale: torch.Tensor | None,
k: torch.Tensor,
k: torch.Tensor | None,
weights: torch.Tensor,
quant_block_size: int,
scale_fmt: str | None,
@@ -695,7 +695,7 @@ def sparse_attn_indexer_fake(
kv_cache: torch.Tensor,
q_quant: torch.Tensor,
q_scale: torch.Tensor | None,
k: torch.Tensor,
k: torch.Tensor | None,
weights: torch.Tensor,
quant_block_size: int,
scale_fmt: str | None,
@@ -781,7 +781,7 @@ class SparseAttnIndexer(CustomOp):
self,
hidden_states: torch.Tensor,
q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
k: torch.Tensor,
k: torch.Tensor | None,
weights: torch.Tensor,
):
if current_platform.is_cuda() or current_platform.is_xpu():
@@ -798,7 +798,7 @@ class SparseAttnIndexer(CustomOp):
self,
hidden_states: torch.Tensor,
q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
k: torch.Tensor,
k: torch.Tensor | None,
weights: torch.Tensor,
):
# FP8 path: single tensor (per-token scale is folded into `weights`).
@@ -835,7 +835,7 @@ class SparseAttnIndexer(CustomOp):
self,
hidden_states: torch.Tensor,
q_fp8: torch.Tensor,
k: torch.Tensor,
k: torch.Tensor | None,
weights: torch.Tensor,
):
return self.forward_cuda(hidden_states, q_fp8, k, weights)
@@ -844,7 +844,7 @@ class SparseAttnIndexer(CustomOp):
self,
hidden_states: torch.Tensor,
q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
k: torch.Tensor,
k: torch.Tensor | None,
weights: torch.Tensor,
):
assert not self.use_fp4_cache, "AMD platform doesn't support fp4 cache yet"