Config patches doc + compress_ratios runtime patch in serve script

scripts/serve_vllm.py

@@ -14,6 +14,47 @@ Or in the background:
 import subprocess
 import sys
 
+# ── Patch: Add compress_ratios to DeepseekV4Config ──────────────────────────
+# transformers 5.8.0 renamed compress_ratios → compress_rates (dict format).
+# vllm 0.20.2 still expects compress_ratios as a list indexed by layer_id.
+# We patch the Config class to expose compress_ratios as a property that
+# converts the new dict format back to the list format vllm expects.
+import transformers
+try:
+    from transformers import DeepseekV4Config
+
+    _orig_init = DeepseekV4Config.__init__
+
+    def _patched_init(self, *args, **kwargs):
+        _orig_init(self, *args, **kwargs)
+        # If compress_ratios already exists as a list, leave it alone
+        if hasattr(self, 'compress_ratios') and isinstance(self.compress_ratios, list):
+            return
+        # Convert compress_rates dict → compress_ratios list
+        if hasattr(self, 'compress_rates') and isinstance(self.compress_rates, dict):
+            rates = self.compress_rates
+            # Build per-layer list from the dict schema
+            # V4 pattern: layers 0-1=128, then alternating 4/128, last=0
+            n_layers = getattr(self, 'num_hidden_layers', 61)
+            cr = rates.get('compressed_sparse_attention', 4)
+            hr = rates.get('heavily_compressed_attention', 128)
+            ratios = []
+            for i in range(n_layers):
+                if i < 2:
+                    ratios.append(hr)
+                elif i == n_layers - 1:
+                    ratios.append(0)
+                else:
+                    ratios.append(cr if i % 2 == 0 else hr)
+            self.compress_ratios = ratios
+        elif hasattr(self, 'compress_rates') and isinstance(self.compress_rates, list):
+            self.compress_ratios = self.compress_rates
+
+    DeepseekV4Config.__init__ = _patched_init
+    print("✓ Patched DeepseekV4Config.__init__ to add compress_ratios")
+except ImportError:
+    print("⚠ DeepseekV4Config not found, skipping compress_ratios patch")
+
 MODEL = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
 
 # These flags are critical for V4 — do not change without understanding why: