[1/2] Refactor DeepGeem requant for FP8 Linear on Blackwell (#13601)

Co-authored-by: fy1214
2025-11-23 16:07:56 -08:00
parent 9054e844ea
commit 4683e244fe
4 changed files with 39 additions and 79 deletions
@@ -220,7 +220,6 @@ class DeepseekV3ForCausalLMNextN(DeepseekV3ForCausalLM):
            use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
        )
        self.logits_processor = LogitsProcessor(config)
-        self._executed_weight_requant_ue8m0 = False

    @torch.no_grad()
    def forward(