Ovis padding

.jenkins/lm-eval-harness/configs/models-ovis.txt

@@ -0,0 +1 @@
+ovis2_5-9b.yaml

.jenkins/lm-eval-harness/configs/ovis2_5-9b.yaml

@@ -0,0 +1,12 @@
+model_name: "/mnt/weka/data/llm/aidc-ai/ovis2.5-9b"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.700
+  - name: "exact_match,flexible-extract"
+    value: 0.700
+limit: 256
+num_fewshot: 8
+dtype: "bfloat16"
+trust_remote_code: True

.jenkins/test_config.yaml

@@ -5,54 +5,59 @@ stages:
       - name: v0_gsm8k_small_g3_tp1_part1
         flavor: g3
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small.txt -t 1
       - name: v0_gsm8k_small_g3_tp1_part2
         flavor: g3
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small-2.txt -t 1
       - name: v0_gsm8k_small_g3_tp1_part3
         flavor: g3
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small-3.txt -t 1
       - name: v0_gsm8k_small_g3_tp2
         flavor: g3.s
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small.txt -t 2
       - name: v0_gsm8k_small_g2_tp1
         flavor: g2
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small.txt -t 1
       - name: v0_gsm8k_small_g2_tp2
         flavor: g2.s
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small.txt -t 2
       - name: v0_gsm8k_g2_deepseek-v2-lite_tp1
         flavor: g3
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-deepseek.txt -t 1
       - name: v0_gsm8k_g3_gemma3_tp1
         flavor: g3.s
         command: >-
           export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-gemma.txt -t 1
+      - name: v0_gsm8k_g3_ovis2_5_tp1
+        flavor: g3.s
+        command: >-
+          export PT_HPU_LAZY_MODE=1 && export VLLM_SKIP_WARMUP=true &&
+          cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-ovis.txt -t 1
   - name: test_gsm8k_small_models_apc
     steps:
       - name: gsm8k_small_g3_tp1_apc
         flavor: g3
         command: >-
-          export VLLM_CONTIGUOUS_PA=false && 
+          export VLLM_CONTIGUOUS_PA=false &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small.txt -t 1 -a
       - name: gsm8k_small_g2_tp1_apc
         flavor: g2
         command: >-
-          export VLLM_CONTIGUOUS_PA=false && 
+          export VLLM_CONTIGUOUS_PA=false &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-small.txt -t 1 -a
   - name: test_gsm8k_small_models_merged_prefill
     steps:
@@ -66,139 +71,139 @@ stages:
       - name: v0_gsm8k_large_g3_tp2_part1
         flavor: g3.s
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-large.txt -t 2
       - name: v0_gsm8k_large_g3_tp2_part2
         flavor: g3.s
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-large-2.txt -t 2
       - name: v0_gsm8k_large_g2_tp4
         flavor: g2.m
         command: >-
-          export PT_HPU_LAZY_MODE=1 && 
+          export PT_HPU_LAZY_MODE=1 &&
           cd .jenkins/lm-eval-harness && bash run-tests.sh -c configs/models-large.txt -t 4
   - name: test_gsm8k_fp8
     steps:
       - name: gsm8k_small_g3_tp1_fp8
         flavor: g3
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-fp8-g3-tp1.txt -t 1
       # - name: gsm8k_small_g3_tp2_fp8
       #   flavor: g3.s
       #   command: >-
-      #     cd .jenkins/lm-eval-harness && 
-      #     PT_HPU_LAZY_MODE=1 
+      #     cd .jenkins/lm-eval-harness &&
+      #     PT_HPU_LAZY_MODE=1
       #     bash run-tests.sh -c configs/models-fp8.txt -t 2
   - name: test_gsm8k_fp8_bypass_inc
     steps:
       - name: gsm8k_fp8_llama4_scout_g3_tp2_compressed_tensor
         flavor: g3.s
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-fp8-compressedtensor.txt -t 2
       - name: gsm8k_fp8_qwen3_30B_g3_tp1_block_scale_dynamic
         flavor: g3
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-fp8-blockfp8.txt -t 1
       - name: gsm8k_fp8_qwen3_30B_g3_tp1_block_scale_dequant
         flavor: g3
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 VLLM_HPU_FORCE_CHANNEL_FP8=0 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1 VLLM_HPU_FORCE_CHANNEL_FP8=0
           bash run-tests.sh -c configs/models-fp8-blockfp8.txt -t 1
   - name: test_gsm8k_mss
     steps:
       - name: gsm8k_small_g3_tp1_mss
         flavor: g3
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 1
       - name: gsm8k_small_g2_tp1_mss
         flavor: g2
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 1
       - name: gsm8k_small_g3_tp2_mss
         flavor: g3.s
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 2
       - name: gsm8k_small_g2_tp2_mss
         flavor: g2.s
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 2
       - name: gsm8k_small_g2_tp1_spec_decode
         flavor: g2
         command: >-
-          cd .jenkins/lm-eval-harness && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/lm-eval-harness &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 1
   - name: test_gsm8k_spec_decode
     steps:
       # - name: gsm8k_small_g2_tp1_mlp_spec_decode
       #   flavor: g2
       #   command: >-
-      #     PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True 
+      #     PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True
       #     pytest -v tests/spec_decode/e2e/test_mlp_correctness.py::test_mlp_e2e_greedy_correctness
       - name: gsm8k_small_g2_tp1_medusa_spec_decode
         flavor: g2
         command: >-
-          PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True 
+          PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True
           pytest -v tests/spec_decode/e2e/test_medusa_correctness.py::test_medusa_e2e_greedy_correctness
       - name: gsm8k_small_g2_tp1_eagle_spec_decode
         flavor: g2
         command: >-
-          PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True 
+          PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True
           pytest -v tests/spec_decode/e2e/test_eagle_correctness.py::test_eagle_e2e_greedy_correctness
   #TODO(kwisniewski98) temporary disable test, until model specific for Gaudi2 is uploaded to test infrastructure
   # - name: test_deepseek_mtp
   #   steps:
   #     - name: test_deepseek_mtp_correctness
   #       flavor: g3
   #       command: >-
-  #         PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True 
+  #         PT_HPU_LAZY_MODE=1 VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True
   #         pytest -v tests/spec_decode/e2e/test_mtp_correctness.py::test_mtp_e2e_greedy_correctness
   - name: tests_lora
     steps:
       - name: test_llama_lora
         flavor: g2
         command: >-
-          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true 
+          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true
           pytest -v tests/lora/test_llama_hpu.py::test_llama_lora_1x
       - name: test_multilora
         flavor: g2
         command: >-
-          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true 
+          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true
           pytest -v tests/lora/test_multilora_hpu.py::test_llama_multilora_1x
       # - name: test_long_context
       #   flavor: g2
       #   command: >-
-      #     PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true 
+      #     PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true
       #     pytest -v tests/lora/test_long_context_hpu.py::test_quality
   - name: tests_multimodal
     steps:
       - name: multimodal_small_g3_tp1
         flavor: g3
         command: >-
-          cd .jenkins/vision && 
-          PT_HPU_LAZY_MODE=1 
+          cd .jenkins/vision &&
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-small.txt -t 1
       - name: multimodal_small_g3_tp2
         flavor: g3.s
         command: >-
           cd .jenkins/vision &&
-          PT_HPU_LAZY_MODE=1 
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-small.txt -t 2
       - name: multimodal_qwen_tp1
         flavor: g3.s
@@ -210,13 +215,13 @@ stages:
         flavor: g3
         command: >-
           cd .jenkins/vision &&
-          PT_HPU_LAZY_MODE=1 
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 1
       - name: multimodal_small_g3_tp2_mss
         flavor: g3.s
         command: >-
           cd .jenkins/vision &&
-          PT_HPU_LAZY_MODE=1 
+          PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-mss.txt -t 2
       - name: multimodal_llama4_scout_g3_tp2_ep
         flavor: g3.s
@@ -230,26 +235,32 @@ stages:
           cd .jenkins/vision &&
           PT_HPU_LAZY_MODE=1
           bash run-tests.sh -c configs/models-gemma.txt -t 1
+      - name: multimodal_ovis2_5_g3_tp1_ep
+        flavor: g3.s
+        command: >-
+          cd .jenkins/vision &&
+          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true
+          bash run-tests.sh -c configs/models-ovis.txt -t 1
   - name: tests_int4_quantization
     steps:
       - name: test_awq
         flavor: g2
         command: >-
-          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true 
+          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true
           pytest -v tests/quantization/test_awq.py::test_awq
       - name: test_gptq
         flavor: g2
         command: >-
-          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true 
+          PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true
           pytest -v tests/quantization/test_gptq.py::test_gptq
   - name: tests_guided_decode
     steps:
     - name: test_lazy_outlines
       flavor: g2
       command: >-
         pip install -e tests/vllm_test_utils &&
-        export VLLM_SKIP_WARMUP=true && PT_HPU_LAZY_MODE=1 
-        pytest -v tests/entrypoints/llm/test_lazy_outlines.py -s -vvv --log-cli-level=INFO 
+        export VLLM_SKIP_WARMUP=true && PT_HPU_LAZY_MODE=1
+        pytest -v tests/entrypoints/llm/test_lazy_outlines.py -s -vvv --log-cli-level=INFO
     # - name: test_guided_generate
     #   flavor: g2
     #   command: >-

.jenkins/vision/configs/models-ovis.txt

@@ -0,0 +1 @@
+ovis2_5-9b.yaml

.jenkins/vision/configs/ovis2_5-9b.yaml

@@ -0,0 +1,7 @@
+model_name: "/mnt/weka/data/llm/aidc-ai/ovis2.5-9b"
+dtype: "bfloat16"
+max_model_len: 32768
+max_num_seqs: 32
+num_prompts: 4
+limit_mm_per_prompt_image: 5
+trust_remote_code: True

.jenkins/vision/test_enc_dec_model.py

@@ -24,6 +24,7 @@ def fail_on_exit():
 def launch_enc_dec_model(config, question, images):
     model_name = config.get('model_name')
     dtype = config.get('dtype', 'bfloat16')
+    trust_remote_code = config.get('trust_remote_code', False)
     max_num_seqs = config.get('max_num_seqs', 128)
     max_model_len = config.get('max_model_len', 4096)
     enforce_eager = config.get('enforce_eager', False)
@@ -41,6 +42,7 @@ def launch_enc_dec_model(config, question, images):
         enable_expert_parallel=enable_expert_parallel,
         enforce_eager=enforce_eager,
         limit_mm_per_prompt={"image": limit_mm_per_prompt_image},
+        trust_remote_code=trust_remote_code,
     )
 
     tokenizer = AutoTokenizer.from_pretrained(model_name)

examples/offline_inference/vision_language_multi_image.py

@@ -460,6 +460,36 @@ def load_ovis(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+# ovis2_5
+def load_ovis2_5(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "AIDC-AI/Ovis2.5-2B"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        max_num_seqs=2,
+        trust_remote_code=True,
+        dtype="half",
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholders = "\n".join(
+        f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1)
+    )
+    messages = [{"role": "user", "content": f"{placeholders}\n{question}"}]
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+    prompt = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+    )
+
+
 def load_pixtral_hf(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "mistral-community/pixtral-12b"
 
@@ -742,6 +772,7 @@ def load_tarsier(question: str, image_urls: list[str]) -> ModelRequestData:
     "mllama": load_mllama,
     "NVLM_D": load_nvlm_d,
     "ovis": load_ovis,
+    "ovis2_5": load_ovis2_5,
     "phi3_v": load_phi3v,
     "phi4_mm": load_phi4mm,
     "pixtral_hf": load_pixtral_hf,