fix: use tqdm for progress bars — single line, live updating

Replaces manual bar printing with tqdm. Overwrites the same line
instead of spewing one line per update.

vllm/patches/deepseek_v4.py

@@ -428,12 +428,8 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         l1_fp4, l1_sf, l1_gs = [], [], []
         l2_fp4, l2_sf, l2_gs = [], [], []
 
-        n_exp = self.num_local_experts
-        for e in range(n_exp):
-            if e % max(1, n_exp // 10) == 0 or e == n_exp - 1:
-                pct = (e + 1) * 100 // n_exp
-                bar = "█" * (pct // 5) + "░" * (20 - pct // 5)
-                print(f"  NVFP4 experts [{bar}] {pct:3d}% ({e+1}/{n_exp})", flush=True)
+        from tqdm import tqdm
+        for e in tqdm(range(self.num_local_experts), desc="  NVFP4 experts", unit="exp"):
             # ── L1: gate + up (fused) ──
             gate_w = self.w13_weight.data[e, :self.intermediate_size]  # (intermediate, hidden//2) uint8
             up_w = self.w13_weight.data[e, self.intermediate_size:]    # (intermediate, hidden//2) uint8
@@ -1626,12 +1622,8 @@ class DeepseekV4Model(nn.Module):
         # Build shard index once for compressor reconstruction (avoids N×M full-shard loads)
         _shard_index = self._build_shard_index("/model") if os.path.isdir("/model") else None
 
-        n_layers = len(self.layers)
-        for layer_idx, layer in enumerate(self.layers):
-            if layer_idx % max(1, n_layers // 10) == 0 or layer_idx == n_layers - 1:
-                pct = (layer_idx + 1) * 100 // n_layers
-                bar = "█" * (pct // 5) + "░" * (20 - pct // 5)
-                print(f"  NVFP4 convert [{bar}] {pct:3d}% (layer {layer_idx+1}/{n_layers})", flush=True)
+        from tqdm import tqdm
+        for layer_idx, layer in tqdm(enumerate(self.layers), total=len(self.layers), desc="  NVFP4 convert", unit="layer"):
             attn = layer.attn
             
             # FP8 conversion: only wo_a