Update test-pipeline.yaml

.buildkite/test-pipeline.yaml

@@ -420,7 +420,7 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader

docs/dev-docker/README.md

@@ -291,7 +291,8 @@ python3 /app/vllm/benchmarks/benchmark_throughput.py \
     --num-prompts $PROMPTS \
     --max-num-seqs $MAX_NUM_SEQS
 ```
-For FP16 models, remove `--kv-cache-dtype fp8`. 
+
+For FP16 models, remove `--kv-cache-dtype fp8`.
 
 When measuring models with long context lengths, performance may improve by setting `--max-model-len` to a smaller value (8192 in this example).  It is important, however, to ensure that the `--max-model-len` is at least as large as the IN + OUT token counts.
 
@@ -325,6 +326,7 @@ vllm serve amd/Llama-3.1-70B-Instruct-FP8-KV \
     --gpu-memory-utilization 0.99 \
     --num_scheduler-steps 10
 ```
+
 For FP16 models, remove `--kv-cache-dtype fp8`. Change port (for example --port 8005) if port=8000 is currently being used by other processes.
 
 Run client in a separate terminal. Use port_id from previous step else port-id=8000.