[DSV4] Add BF16 and MXFP8 A2A support for flashinfer a2a one sided (#40960)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Signed-off-by: Zijing Liu <liuzijing2014@gmail.com>
Co-authored-by: Zijing Liu <liuzijing2014@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-04-30 15:33:12 -07:00
committed by GitHub
co-authored by Zijing Liu Claude Opus 4.7
parent 526927be94
commit b4806c8ee1
12 changed files with 107 additions and 68 deletions
@@ -577,6 +577,8 @@ class FlashInferNVLinkOneSidedManager(All2AllManagerBase):
top_k: int,
num_experts: int,
hidden_size: int,
dispatch_dtype_bytes_per_elem: int = 0,
dispatch_scale_bytes_per_token: int = 0,
):
"""Initialize the MoeAlltoAll workspace."""
if self.initialized:
@@ -607,9 +609,13 @@ class FlashInferNVLinkOneSidedManager(All2AllManagerBase):
ep_config = MnnvlConfig(
comm_backend=CustomCommunicator(self.cpu_group),
)
if dispatch_dtype_bytes_per_elem == 0:
hidden_bytes = hidden_size // 2
else:
hidden_bytes = hidden_size * dispatch_dtype_bytes_per_elem
total_dispatch_payload_size_per_token = (
hidden_size // 2 # nvfp4 hidden states
+ hidden_size // 16 # fp8 scaling factors
hidden_bytes
+ dispatch_scale_bytes_per_token
+ top_k * 4 # int32 topks ids
+ top_k * 4 # float32 topk weights
)