disable moe_calibrate_all_experts for Autoround example.

examples/autoround/quantization_w8a8_fp8/llama4_dynamic_quant_example.py

@@ -48,6 +48,9 @@
     max_seq_length=MAX_SEQUENCE_LENGTH,
     num_calibration_samples=NUM_CALIBRATION_SAMPLES,
     shuffle_calibration_samples=False,
+    # Avoid calibrating all experts in this single-GPU example path to reduce
+    # OOM risk when llmcompressor has already linearized MoE experts.
+    moe_calibrate_all_experts=False,
 )
 
 # Confirm generations of the quantized model look sane.

examples/autoround/quantization_w8a8_fp8/llama4_static_quant_example.py

@@ -48,6 +48,9 @@
     max_seq_length=MAX_SEQUENCE_LENGTH,
     num_calibration_samples=NUM_CALIBRATION_SAMPLES,
     shuffle_calibration_samples=False,
+    # Avoid calibrating all experts in this single-GPU example path to reduce
+    # OOM risk when llmcompressor has already linearized MoE experts.
+    moe_calibrate_all_experts=False,
 )
 
 # Confirm generations of the quantized model look sane.

examples/autoround/quantization_w8a8_mxfp8/autoround_example.py

@@ -77,6 +77,10 @@
     num_calibration_samples=NUM_CALIBRATION_SAMPLES,
     # disable shuffling to get slightly better mmlu score
     shuffle_calibration_samples=False,
+    # For Qwen3-30B, 128 samples, 200 iters, expert routing coverage is
+    # sufficient in practice. Enabling all experts can OOM on a single 80GB GPU,
+    # so keep this False for the single-GPU example path.
+    moe_calibrate_all_experts=False,
 )
 
 # Confirm generations of the quantized model look sane.