glm-4v add input process

examples/offline_inference_vision_language.py

@@ -11,6 +11,9 @@
 from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
 from vllm.utils import FlexibleArgumentParser
+import os
+
+os.environ['CUDA_VISIBLE_DEVICES'] = '3'
 
 
 # LLaVA-1.5
@@ -196,6 +199,23 @@ def run_qwen2_vl(question):
     return llm, prompt, stop_token_ids
 
 
+def run_glm4v(question):
+    model_name = "THUDM/glm-4v-9b"
+    model_name = "/workspace/siaowei/model/glm-4v-9b"
+
+    llm = LLM(
+        model=model_name,
+        tensor_parallel_size=1,
+        max_model_len=8192,
+        trust_remote_code=True,
+        # gpu_memory_utilization=0.5,
+        enforce_eager=True)
+    # prompt = f"[gMASK]<sop><|user|>{question}<|begin_of_image|><|endoftext|><|end_of_image|><|assistant|>"
+    prompt = question
+    stop_token_ids = [151329, 151336, 151338]
+    return llm, prompt, stop_token_ids
+
+
 model_example_map = {
     "llava": run_llava,
     "llava-next": run_llava_next,
@@ -209,6 +229,7 @@ def run_qwen2_vl(question):
     "internvl_chat": run_internvl,
     "qwen_vl": run_qwen_vl,
     "qwen2_vl": run_qwen2_vl,
+    "glm4v": run_glm4v,
 }
 
 

tests/models/decoder_only/vision_language/test_glm4.py

@@ -0,0 +1,121 @@
+# tests/models/decoder_only/vision_language/test_glm4v.py
+from typing import List, Optional, Tuple, Type, Dict
+
+import pytest
+import os
+from vllm.sequence import SampleLogprobs
+from vllm.multimodal.utils import rescale_image_size
+from vllm.utils import is_cpu
+from transformers import AutoConfig, AutoTokenizer
+
+
+from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner)
+from ...utils import check_logprobs_close
+
+HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
+    "stop_sign":
+    "What's the content of the image?",
+    "cherry_blossom":
+    "What is the season?",
+})
+os.environ['CUDA_VISIBLE_DEVICES'] = '3'
+
+models = ["/workspace/siaowei/model/glm-4v-9b"]
+
+target_dtype = "bfloat16"
+if is_cpu():
+    target_dtype = "bfloat16"
+
+def run_test(
+    hf_runner: Type[HfRunner],
+    vllm_runner: Type[VllmRunner],
+    inputs: List[Tuple[List[str], PromptImageInput]],
+    model: str,
+    *,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+    mm_limit: int,
+    tensor_parallel_size: int,
+    distributed_executor_backend: Optional[str] = None,
+):  
+    # max_model_len should be greater than image_feature_size
+    with vllm_runner(model,
+                     max_model_len=4096,
+                     max_num_seqs=1,
+                     dtype=dtype,
+                     limit_mm_per_prompt={"image": mm_limit},
+                     tensor_parallel_size=tensor_parallel_size,
+                     distributed_executor_backend=distributed_executor_backend,
+                    #  gpu_memory_utilization=0.9,
+                     enforce_eager=True) as vllm_model:
+        # tokenizer = vllm_model.model.get_tokenizer()
+        # pass
+        stop_token_ids = [151329, 151336, 151338]
+        vllm_outputs_per_image = [
+                    vllm_model.generate_greedy_logprobs(prompts,
+                                                        max_tokens,
+                                                        num_logprobs=num_logprobs,
+                                                        images=images,
+                                                        stop_token_ids=stop_token_ids)
+                    for prompts, images in inputs
+                ]
+    with hf_runner(model, dtype=dtype) as hf_model:
+        eos_token_id = hf_model.tokenizer.eos_token_id
+        hf_outputs_per_image = [
+            hf_model.generate_greedy_logprobs_limit(prompts,
+                                                    max_tokens,
+                                                    num_logprobs=num_logprobs,
+                                                    images=images,
+                                                    eos_token_id=eos_token_id
+                                                    # tokenizer=tokenizer       
+                                                    )
+            for prompts, images in inputs
+        ]
+
+    for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
+                                        vllm_outputs_per_image):
+
+        check_logprobs_close(
+                outputs_0_lst=hf_outputs,
+                outputs_1_lst=vllm_outputs,
+                name_0="hf",
+                name_1="vllm",
+        )
+
+@pytest.mark.parametrize("model", models)
+@pytest.mark.parametrize(
+    "size_factors",
+    [
+        # No image
+        [],
+        # Single-scale
+        # [1.0],
+        # Single-scale, batched
+        # [1.0, 1.0, 1.0],
+        # Multi-scale
+        # [0.25, 0.5, 1.0],
+    ],
+)
+@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize("max_tokens", [128])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
+                dtype: str, max_tokens: int, num_logprobs: int) -> None:
+    images = [asset.pil_image for asset in image_assets]
+
+    inputs_per_image = [(
+        [prompt for _ in size_factors],
+        [rescale_image_size(image, factor) for factor in size_factors],
+    ) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
+    run_test(
+        hf_runner,
+        vllm_runner,
+        inputs_per_image,
+        model,
+        dtype=dtype,
+        max_tokens=max_tokens,
+        num_logprobs=num_logprobs,
+        mm_limit=1,
+        tensor_parallel_size=1,
+    )