Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput) (#7667)

This commit is contained in:
Trevor Morris
2025-08-15 22:08:11 -07:00
committed by GitHub
parent 87dab54824
commit eff4eb3fdd
16 changed files with 360 additions and 52 deletions

View File

@@ -191,7 +191,11 @@ class LogitsMetadata:
else:
self.global_dp_buffer_len = self.global_dp_buffer_len
set_dp_buffer_len(self.global_dp_buffer_len, self.dp_local_num_tokens)
set_dp_buffer_len(
self.global_dp_buffer_len,
self.dp_local_num_tokens,
self.global_num_tokens_for_logprob_cpu,
)
class LogitsProcessor(nn.Module):