Lightoncr 2 1b

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

@@ -48,6 +48,14 @@ smoke-test:
       fleet: "x86-g6exl-runner"
       extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16 --trust-remote-code --mamba_ssm_cache_dtype float32"
 
+    # LightOnOCR-2-1B: efficient 1B VLM for document OCR (LightOnOcrForConditionalGeneration).
+    # Mistral3-based, bf16 (~2GB). Requires multimodal flags for image input.
+    # Fits a single L4 (g6.xl, 24GB) easily at tp=1.
+    - name: "lighton-ocr-2-1b"
+      s3_model: "lightonocr-2-1b.tar.gz"
+      fleet: "x86-g6xl-runner"
+      extra_args: "--tensor-parallel-size 1 --max-model-len 4096 --dtype bfloat16 --no-enable-prefix-caching"
+
     - name: "qwen3-embedding-0.6b"
       s3_model: "qwen3-embedding-0.6b.tar.gz"
       fleet: "x86-g6xl-runner"

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

@@ -28,6 +28,16 @@ benchmark:
       test_script: "vllm_embedding_benchmark_test.sh"
       min_rps: 3
 
+    # LightOnOCR-2-1B: 1B VLM for document OCR (LightOnOcrForConditionalGeneration).
+    # bf16 ~2GB, fits L40S easily. Thresholds are provisional — calibrate after first run.
+    - name: "lighton-ocr-2-1b"
+      s3_model: "lightonocr-2-1b.tar.gz"
+      runner_label: "gpu-l40s-1gpu-runners"
+      extra_args: "--dtype bfloat16 --max-model-len 4096"
+      min_throughput: 30
+      min_rps: 1
+      benchmark_profiles: "baseline,high_concurrency"
+
     # Realtime model — benchmark via /v1/audio/transcriptions; it does not
     # support chat audio_url (empty MM embeddings crash the engine).
     - name: "voxtral-mini-4b"

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

@@ -70,6 +70,32 @@ sagemaker:
             temperature: 0
         validate: "json_field:choices.0.message.content"
 
+  # LightOnOCR-2-1B — efficient 1B VLM for document OCR, bf16 (~2GB).
+  # LightOnOcrForConditionalGeneration (Mistral3-based). Text-only chat completion
+  # validates serving; full OCR with image input requires multimodal request support.
+  - name: "lighton-ocr-2-1b"
+    s3_model: "lightonocr-2-1b.tar.gz"
+    instance_type: "ml.g6.xlarge"
+    required_image_pattern:
+      os_version: "amzn2023"
+    env:
+      SM_VLLM_MODEL: "/opt/ml/model"
+      SM_VLLM_MAX_MODEL_LEN: "4096"
+      SM_VLLM_DTYPE: "bfloat16"
+    test_cases:
+      - name: "chat-completion"
+        route: "/v1/chat/completions"
+        content_type: "application/json"
+        request:
+          body:
+            model: "/opt/ml/model"
+            messages:
+              - role: "user"
+                content: "What is optical character recognition?"
+            max_tokens: 256
+            temperature: 0
+        validate: "json_field:choices.0.message.content"
+
   - name: "voxtral-mini-4b"
     s3_model: "voxtral-mini-4b.tar.gz"
     instance_type: "ml.g6.xlarge"