Fix warmup: pass local expert IDs (not global), remove incorrect _warmup_done guard

compute_activation_global_scales expects local IDs (0..num_experts-1),
not global IDs. EP5/EP7 were getting L2 gs=0 because global IDs (240+,
336+) didn't match expert_id_range (0..47), so no tokens matched any
expert → L1 GEMM got zero inputs → L2 gs=0 → NaN/crash.

Also removed _warmup_done guard since each layer needs its own warmup
(different weights, different gs values).

vllm/patches/deepseek_v4.py

@@ -551,19 +551,14 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         # the actual amax and compute correct gs values.
         self._warmup_activation_global_scales()
 
-    _warmup_done: bool = False
-
     def _warmup_activation_global_scales(self) -> None:
         """Run a warmup forward pass to compute correct activation global scales.
         
-        Called once during finalize_weights, before cudagraph capture.
+        Called once per layer during finalize_weights, before cudagraph capture.
         Uses quantize_to_nvfp4 (which calls .max()) to get the exact gs
         from real activation magnitudes, then stores them for use by
         quantize_activation_nvfp4 (no .max(), cudagraph-safe).
         """
-        if self._warmup_done:
-            return
-        self._warmup_done = True
         import torch
         runner = self._cutedsl_runner
         device = runner.device
@@ -574,11 +569,12 @@ class DeepseekV4MegaMoEExperts(nn.Module):
             # Sample hidden states: typical BF16 activations have amax ~1-10
             hidden_states = torch.randn(num_tokens, runner.hidden_size,
                                         dtype=torch.bfloat16, device=device)
-            # Assign all tokens to the first few experts evenly
+            # Assign all tokens to local experts (0..num_local_experts-1)
+            # compute_activation_global_scales expects local IDs, not global
             topk_ids = torch.zeros(num_tokens, top_k, dtype=torch.int64, device=device)
             for i in range(num_tokens):
                 for j in range(top_k):
-                    topk_ids[i, j] = (runner.experts_start_idx + j) % (runner.experts_start_idx + runner.num_experts)
+                    topk_ids[i, j] = j % runner.num_experts
             topk_weights = torch.ones(num_tokens, top_k, dtype=torch.float32, device=device) / top_k
             
             runner.compute_activation_global_scales(