clean up

docker/docker-compose.yml

@@ -7,10 +7,16 @@ services:
       - "8000:8000"
     environment:
       - OMP_NUM_THREADS=128
+      - CUDA_LAUNCH_BLOCKING=0
+      - TORCH_SHOW_CPP_STACKTRACES=0
       - MEGA_MOE_DEBUG=0
       - MEGA_MOE_STATIC=0
+      - NVFP4_DEBUG=0
+      - NVFP4_DEBUG_SYNC=0
+      - SKIP_ATTENTION=0
       - MEGA_MOE_USE_CUTLASS=1
       - DG_JIT_DEBUG=0
+      - DEEP_GEMM_JIT_DEBUG=0
     command:
       - /model
       - --trust-remote-code

docker/patches/deepseek_v4.py

@@ -1586,10 +1586,11 @@ class DeepseekV4Model(nn.Module):
                 try:
                     weight_loader(param, loaded_weight, shard_id)
                 except (AssertionError, ValueError, RuntimeError) as e:
-                    print(f'[DEBUG-STACK] FAILED: name={name} shard_id={shard_id} '
-                          f'param.shape={param.shape} param.dtype={param.data.dtype} '
-                          f'loaded.shape={loaded_weight.shape} loaded.dtype={loaded_weight.dtype} err={e}')
-                    raise
+                    raise RuntimeError(
+                        f'Weight load failed: name={name} shard_id={shard_id} '
+                        f'param.shape={param.shape} param.dtype={param.data.dtype} '
+                        f'loaded.shape={loaded_weight.shape} loaded.dtype={loaded_weight.dtype}'
+                    ) from e
                 loaded_params.add(name)
                 break
             else:
@@ -1866,7 +1867,7 @@ class DeepseekV4Model(nn.Module):
         
         total_fp8 = fp8_converted + fp8_from_bf16
         total_bf16 = bf16_converted + compressor_converted
-        if total_fp8 > 0 or total_bf16 > 0:
+        if int(os.environ.get('NVFP4_DEBUG', '0')) and (total_fp8 > 0 or total_bf16 > 0):
             print(f"NVFP4 post-load: {fp8_converted} NVFP4->FP8, "
                   f"{fp8_from_bf16} BF16->FP8, "
                   f"{bf16_converted} attn/shared->BF16, "