fix glob to path

Signed-off-by: sirutBuasai <sirutbuasai27@outlook.com>

Author: sirutBuasai

.github/actions/build-image/action.yml

@@ -49,6 +49,8 @@ runs:
         echo "os-version=$(yq '.metadata.os_version' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
         echo "python-version=$(yq '.build.python_version // ""' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
         echo "cuda-version=$(yq '.build.cuda_version // ""' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
+        echo "contributor=$(yq '.metadata.contributor // ""' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
+        echo "transformers-version=$(yq '.build.transformers_version // ""' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
         echo "dockerfile=$(yq '.build.dockerfile' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
         echo "target=$(yq '.build.target // ""' "$CONFIG_FILE")" >> $GITHUB_OUTPUT
 
@@ -152,6 +154,17 @@ runs:
           LABEL_ARGS+=("--label" "com.amazonaws.ml.engines.sagemaker.dlc.job.${JOB_TYPE}=true")
           [[ -n "$LABEL_OS" ]] && LABEL_ARGS+=("--label" "com.amazonaws.ml.engines.sagemaker.dlc.os.${LABEL_OS}=true")
           [[ -n "$LABEL_PYTHON" ]] && LABEL_ARGS+=("--label" "com.amazonaws.ml.engines.sagemaker.dlc.python.${LABEL_PYTHON}=true")
+
+          CONTRIBUTOR="${{ steps.metadata.outputs.contributor }}"
+          if [[ -n "$CONTRIBUTOR" && "$CONTRIBUTOR" != "None" ]]; then
+            LABEL_ARGS+=("--label" "com.amazonaws.ml.engines.sagemaker.dlc.contributor.${CONTRIBUTOR}=true")
+          fi
+
+          TRANSFORMERS_VER="${{ steps.metadata.outputs.transformers-version }}"
+          if [[ -n "$TRANSFORMERS_VER" ]]; then
+            LABEL_TRANSFORMERS=$(echo "${TRANSFORMERS_VER}" | tr '.' '-')
+            LABEL_ARGS+=("--label" "com.amazonaws.ml.engines.sagemaker.dlc.lib.transformers.${LABEL_TRANSFORMERS}=true")
+          fi
         fi
 
         # Target flag (optional)

.github/config/image/huggingface-vllm/sagemaker.yml

@@ -0,0 +1,36 @@
+image:
+  name: "huggingface-vllm-sagemaker"
+  description: "Hugging Face vLLM for SageMaker (layered on the AWS vLLM DLC)"
+
+metadata:
+  framework: "huggingface_vllm"
+  framework_version: "0.22.1"
+  os_version: "ubuntu22.04"
+  customer_type: "sagemaker"
+  platform: "sagemaker"
+  arch_type: "x86"
+  device_type: "gpu"
+  job_type: "general"
+  contributor: "huggingface"
+  prod_image: "huggingface-vllm:server-sagemaker-cuda-v1"
+
+build:
+  dockerfile: "docker/huggingface/vllm/Dockerfile"
+  target: "huggingface-vllm-sagemaker"
+  base_image: "public.ecr.aws/deep-learning-containers/vllm:0.22.1-gpu-py312-cu130-ubuntu22.04-sagemaker-v1"
+  python_version: "3.12"
+  cuda_version: "13.0.2"
+  transformers_version: "5.10.2"
+  huggingface_hub_version: "1.17.0"
+  hf_xet_version: "1.5.0"
+  ffmpeg_tag: "n8.1"
+  dlc_major_version: "1"
+  dlc_minor_version: "0"
+
+release:
+  release: false
+  force_release: false
+  public_registry: false
+  private_registry: true
+  enable_soci: true
+  environment: "production"

.github/config/model-tests/vllm-model-tests.yml

@@ -1,10 +1,13 @@
 # vLLM Model Test Configuration
-# Single config for all model-based test suites.
+# Smoke-test and SageMaker endpoint test suites.
 #
 # smoke-test:  model serving + inference validation
-# benchmark:   throughput and latency with pass/fail thresholds
+# sagemaker:   config-driven real SageMaker endpoint tests
 #
-# Each section has codebuild-fleet and runner-scale-sets sub-keys.
+# Benchmark (throughput/latency) configs were split out into
+# vllm-text-benchmark-tests.yml and vllm-multimodal-benchmark-tests.yml.
+#
+# smoke-test has codebuild-fleet and runner-scale-sets sub-keys.
 # Workflow parsers construct s3_path from s3_prefix + s3_model.
 
 s3_prefix: "s3://dlc-cicd-models/llm-models"
@@ -63,317 +66,6 @@ smoke-test:
 
   runner-scale-sets: []
 
-benchmark:
-  codebuild-fleet:
-    - name: "qwen3-embedding-0.6b"
-      s3_model: "qwen3-embedding-0.6b.tar.gz"
-      fleet: "x86-g6xl-runner"
-      extra_args: "--dtype bfloat16 --max-model-len 8192"
-      test_script: "vllm_embedding_benchmark_test.sh"
-      min_rps: 5
-
-    - name: "qwen3-vl-embedding-2b"
-      s3_model: "qwen3-vl-embedding-2b.tar.gz"
-      fleet: "x86-g6xl-runner"
-      extra_args: "--runner pooling --dtype bfloat16 --max-model-len 2048 --trust-remote-code"
-      test_script: "vllm_embedding_benchmark_test.sh"
-      min_rps: 3
-
-    - name: "qwen3-reranker-4b"
-      s3_model: "qwen3-reranker-4b.tar.gz"
-      fleet: "x86-g6xl-runner"
-      extra_args: "--dtype bfloat16 --max-model-len 10000 --gpu-memory-utilization 0.85"
-      test_script: "vllm_reranker_benchmark_test.sh"
-      min_rps: 20
-
-    - name: "qwen3-asr-1.7b"
-      s3_model: "qwen3-asr-1.7b.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16"
-      test_script: "vllm_asr_benchmark_test.sh"
-      test_fixtures:
-        - "audio/asr_en.wav"
-        - "audio/asr_zh.wav"
-      benchmark_audio_fixture: "asr_en.wav"
-      min_throughput: 30
-      min_rps: 1
-      benchmark_profiles: "baseline,high_concurrency,sustained_load,burst"
-
-    - name: "voxtral-mini-4b"
-      s3_model: "voxtral-mini-4b.tar.gz"
-      fleet: "x86-g6xl-runner"
-      extra_args: "--tokenizer-mode mistral --config-format mistral --load-format mistral --max-model-len 8192 --enforce-eager"
-      test_script: "vllm_asr_benchmark_test.sh"
-      test_fixtures:
-        - "audio/asr_en.wav"
-      benchmark_audio_fixture: "asr_en.wav"
-      min_throughput: 1
-      min_rps: 0.25
-      benchmark_profiles: "baseline,high_concurrency"
-
-    - name: "gpt-oss-20b"
-      s3_model: "gpt-oss-20b.tar.gz"
-      fleet: "x86-g6exl-runner"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 1200
-      min_rps: 5
-
-    - name: "gemma-4-26b-a4b-it"
-      s3_model: "gemma-4-26b-a4b-it.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --dtype bfloat16"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 300
-      min_rps: 2.4
-
-    - name: "gemma-4-31b-it"
-      s3_model: "gemma-4-31b-it.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --dtype bfloat16"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 150
-      min_rps: 1.2
-
-    - name: "gemma-4-e4b-it"
-      s3_model: "gemma-4-e4b-it.tar.gz"
-      fleet: "x86-g6exl-runner"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 680
-      min_rps: 5.3
-
-    - name: "gemma-4-e2b-it"
-      s3_model: "gemma-4-e2b-it.tar.gz"
-      fleet: "x86-g6xl-runner"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 470
-      min_rps: 3.7
-
-    # Pending p5e.48xlarge fleet creation. Fleet name "x86-p5e-runner" is a placeholder.
-    # - name: "minimax-m2.7"
-    #   s3_model: "minimax-m2.7.tar.gz"
-    #   fleet: "x86-p5e-runner"
-    #   extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --dtype auto --trust-remote-code"
-    #   input_len: 512
-    #   output_len: 128
-    #   num_prompts: 64
-    #   batch_size: 4
-    #   min_throughput: 950
-    #   min_rps: 7.4
-
-    # - name: "glm-5.1"
-    #   s3_model: "glm-5.1.tar.gz"
-    #   fleet: "x86-p5e-runner"
-    #   extra_args: "--tensor-parallel-size 8 --max-model-len 4096 --dtype auto --trust-remote-code"
-    #   input_len: 512
-    #   output_len: 128
-    #   num_prompts: 64
-    #   batch_size: 4
-    #   min_throughput: 330
-    #   min_rps: 2.6
-
-    - name: "qwen3.5-9b"
-      s3_model: "qwen3.5-9b.tar.gz"
-      fleet: "x86-g6xl-runner"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096  --enforce-eager"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 20
-      min_rps: 0.15
-
-    - name: "llama-3.3-70b"
-      s3_model: "llama-3.3-70b.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 128
-      num_prompts: 32
-      batch_size: 2
-      min_throughput: 80
-      min_rps: 0.35
-
-    # https://github.com/vllm-project/vllm/issues/32637
-    # transformer version doesn't support this model
-    # https://github.com/vllm-project/vllm/issues/34098
-    # - name: "glm-4.7-flash"
-    #   s3_model: "glm-4.7-flash.tar.gz"
-    #   fleet: "x86-g6xl-runner"
-    #   extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16"
-    #   input_len: 512
-    #   output_len: 128
-    #   num_prompts: 64
-    #   batch_size: 4
-    #   min_throughput: 20
-    #   min_rps: 1
-
-    - name: "qwen3.5-35b-a3b-fp8"
-      s3_model: "qwen3.5-35b-a3b-fp8.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      # https://github.com/vllm-project/vllm/issues/35743 open bug for capturing CUDA graph fails
-      # workaround with --enforce-eager tp=1 fail while tp=4 success
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 80
-      min_rps: 0.35
-
-# A100 is compute capability 8.0 — FP8 requires 8.9+ (H100/L40S).
-# The Marlin fallback uses significantly more memory.
-    - name: "qwen3.5-27b-fp8"
-      s3_model: "qwen3.5-27b-fp8.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 20
-      min_rps: 0.2
-
-    - name: "qwen3-coder-next-fp8"
-      s3_model: "qwen3-coder-next-fp8.tar.gz"
-      fleet: "x86-g6e12xl-runner"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 256
-      num_prompts: 32
-      batch_size: 2
-      min_throughput: 93
-      min_rps: 0.25
-
-  runner-scale-sets:
-    - name: "qwen3-32b"
-      s3_model: "qwen3-32b.tar.gz"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --gpu-memory-utilization 0.85"
-      input_len: 512
-      output_len: 256
-      num_prompts: 32
-      batch_size: 2
-      min_throughput: 1133
-      min_rps: 3
-
-    - name: "qwen3.5-35b-a3b-fp8"
-      s3_model: "qwen3.5-35b-a3b-fp8.tar.gz"
-      # https://github.com/vllm-project/vllm/issues/35743 open bug for capturing CUDA graph fails
-      # workaround with --enforce-eager tp=1 fail while tp=4 success
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 80
-      min_rps: 0.35
-
-    - name: "qwen3.5-27b-fp8"
-      s3_model: "qwen3.5-27b-fp8.tar.gz"
-      # A100 lacks native FP8 — vLLM dequantizes to BF16 at load, doubling weight memory
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --enforce-eager"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 20
-      min_rps: 0.2
-
-    - name: "qwen3-coder-next-fp8"
-      s3_model: "qwen3-coder-next-fp8.tar.gz"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 256
-      num_prompts: 32
-      batch_size: 2
-      min_throughput: 93
-      min_rps: 0.25
-
-    - name: "llama-3.3-70b"
-      s3_model: "llama-3.3-70b.tar.gz"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096"
-      input_len: 512
-      output_len: 128
-      num_prompts: 32
-      batch_size: 2
-      min_throughput: 80
-      min_rps: 0.35
-
-    # --- Qwen 3.5/3.6 new models (thresholds at ~50% of observed) ---
-    - name: "qwen3.5-2b"
-      s3_model: "qwen3.5-2b.tar.gz"
-      runner_label: "gpu-l4-1gpu-runners"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16 --gpu-memory-utilization 0.6"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 5256
-      min_rps: 8.2
-
-    - name: "qwen3.6-27b"
-      s3_model: "qwen3.6-27b.tar.gz"
-      runner_label: "gpu-l40s-4gpu-runners"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --dtype bfloat16 --gpu-memory-utilization 0.8"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 2195
-      min_rps: 3.4
-
-    - name: "qwen3.6-35b-a3b"
-      s3_model: "qwen3.6-35b-a3b.tar.gz"
-      runner_label: "gpu-l40s-4gpu-runners"
-      extra_args: "--tensor-parallel-size 4 --max-model-len 4096 --dtype bfloat16 --gpu-memory-utilization 0.8"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 2654
-      min_rps: 4.1
-
-    - name: "qwen3.5-0.8b"
-      s3_model: "qwen3.5-0.8b.tar.gz"
-      runner_label: "gpu-l4-1gpu-runners"
-      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16 --gpu-memory-utilization 0.6"
-      input_len: 512
-      output_len: 128
-      num_prompts: 64
-      batch_size: 4
-      min_throughput: 5966
-      min_rps: 9.3
-
-# upstream
-# facebook/opt-125m
-# meta-llama/Llama-3.2-1B-Instruct
-# Qwen/Qwen3-0.6B
-# fixie-ai/ultravox-v0_5-llama-3_2-1b
-# llava-hf/llava-1.5-7b-hf
-# microsoft/Phi-3.5-vision-instruct
-# openai/whisper-large-v3-turbo
-# jason9693/Qwen2.5-1.5B-apeach
-# intfloat/e5-small
-# BAAI/bge-reranker-v2-m3
-# meta-llama/Llama-3.1-8B-Instruct
-
 # --- SageMaker Endpoint Tests ---
 # Config-driven real endpoint tests. Each entry deploys a model from S3,
 # sends requests to the configured route, and validates the response.