From 108ff075695139c5710782887a717e31139f3dd7 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Fri, 15 May 2026 03:01:05 +0000 Subject: [PATCH] debug: remove one-shot gate from logit dump, log every forward --- vllm/patches/deepseek_v4.py | 20 +++++++++----------- 1 file changed, 9 insertions(+), 11 deletions(-) diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py index a7d2fe46..e5a046db 100644 --- a/vllm/patches/deepseek_v4.py +++ b/vllm/patches/deepseek_v4.py @@ -2184,17 +2184,15 @@ class DeepseekV4ForCausalLM(nn.Module): hidden_states: torch.Tensor, ) -> torch.Tensor | None: logits = self.logits_processor(self.lm_head, hidden_states) - if not getattr(self, '_logit_dumped', False): - self._logit_dumped = True - last = logits[-1].float() - top_vals, top_idx = last.topk(10) - print(f"[LOGITS] shape={tuple(logits.shape)} " - f"max={last.max().item():.4e} min={last.min().item():.4e} " - f"mean={last.mean().item():.4e} std={last.std().item():.4e} " - f"finite_frac={torch.isfinite(last).float().mean().item():.4f}") - print(f"[LOGITS] top10 ids: {top_idx.tolist()}") - print(f"[LOGITS] top10 vals: {[f'{v:.3f}' for v in top_vals.tolist()]}") - print(f"[LOGITS] gap top1-top10: {(top_vals[0] - top_vals[-1]).item():.3f}") + last = logits[-1].float() + top_vals, top_idx = last.topk(10) + print(f"[LOGITS] shape={tuple(logits.shape)} " + f"max={last.max().item():.4e} min={last.min().item():.4e} " + f"mean={last.mean().item():.4e} std={last.std().item():.4e} " + f"finite_frac={torch.isfinite(last).float().mean().item():.4f}") + print(f"[LOGITS] top10 ids: {top_idx.tolist()}") + print(f"[LOGITS] top10 vals: {[f'{v:.3f}' for v in top_vals.tolist()]}") + print(f"[LOGITS] gap top1-top10: {(top_vals[0] - top_vals[-1]).item():.3f}") return logits def forward(