dang clonkurs

docker-compose.yml

@@ -15,7 +15,8 @@ services:
       - --trust-remote-code
       - --enable-expert-parallel
       - --tensor-parallel-size=8
-      - --enforce-eager
+      - --compilation-config
+      - '{"cudagraph_mode":"FULL_AND_PIECEWISE", "custom_ops":["all"]}'
       - --tokenizer-mode=deepseek_v4
       - --host=0.0.0.0
       - --port=8000

vllm/patches/deepseek_v4.py

@@ -2212,7 +2212,7 @@ class DeepseekV4ForCausalLM(nn.Module):
         # AutoWeightsLoader bypasses this logic and would break NVFP4 loading.
         loaded_params = self.model.load_weights(rest)
         loaded_params.add("lm_head.weight")
-        print("  Checkpoint loaded. Transferring weights to GPU & preparing NVFP4...", flush=True)
+        print("  Checkpoint loaded. Preparing NVFP4...", flush=True)
         self.model.finalize_mega_moe_weights()
         self.model._convert_nvfp4_post_load()
         print("  NVFP4 model ready ✓", flush=True)