fix(musa): stabilize network-backed inference tests

Author: hooray03

tests/functional_tests/inference/qwen3/conf/inference/4b_tp2_musa.yaml

@@ -19,6 +19,9 @@ llm:
   tensor_parallel_size: 2
   pipeline_parallel_size: 1
   gpu_memory_utilization: 0.5
+  # KS3-backed checkpoints must be read sequentially. Lazy mmap loading can
+  # trigger slow random reads and native worker crashes on MUSA.
+  safetensors_load_strategy: eager
   max_model_len: 1024
   seed: 1234
   enforce_eager: true

tests/functional_tests/serve/qwen2_5/conf/0.5b_musa.yaml

@@ -31,6 +31,7 @@ experiment:
     FS_PLATFORM: musa
     TORCHDYNAMO_DISABLE: 1
     TORCH_COMPILE_DISABLE: 1
+    USE_FLAGGEMS: "false"
     VLLM_WORKER_MULTIPROC_METHOD: spawn
     MUSA_VISIBLE_DEVICES: "0"
     MTHREADS_VISIBLE_DEVICES: all

tests/functional_tests/serve/qwen2_5/conf/serve/0.5b_musa.yaml

@@ -19,6 +19,8 @@
     tensor_parallel_size: 1
     pipeline_parallel_size: 1
     gpu_memory_utilization: 0.5
+    # Avoid lazy mmap reads from the network-backed model volume.
+    safetensors_load_strategy: eager
     max_model_len: 1024
     max_num_seqs: 16
     enforce_eager: true