chore: bump vllm to 0.7.2 (#890)

**Reason for Change**:

fix this [issue](vllm-project/vllm#8432)

---------

Signed-off-by: jerryzhuang <[email protected]>

charts/kaito/workspace/templates/inference-params.yaml

@@ -16,4 +16,5 @@ data:
       # max-seq-len-to-capture: 8192
       # num-scheduler-steps: 1
       # enable-chunked-prefill: false
-      # see https://docs.vllm.ai/en/stable/models/engine_args.html for more options.
+      # max-model-len: 2048
+      # see https://docs.vllm.ai/en/stable/serving/engine_args.html for more options.

presets/workspace/dependencies/requirements.txt

@@ -1,9 +1,9 @@
 # Dependencies for TFS
 
 # Core Dependencies
-vllm==0.6.3
-transformers == 4.45.0
-torch==2.4.0
+vllm==0.7.2
+transformers == 4.49.0
+torch==2.5.1
 accelerate==1.0.0
 fastapi>=0.111.0,<0.112.0  # Allow patch updates
 pydantic>=2.9

presets/workspace/inference/vllm/inference_api.py

@@ -12,8 +12,8 @@
 import torch
 from vllm.utils import FlexibleArgumentParser
 import vllm.entrypoints.openai.api_server as api_server
-from vllm.entrypoints.openai.serving_engine import LoRAModulePath
-from vllm.engine.llm_engine import (LLMEngine, EngineArgs, EngineConfig)
+from vllm.entrypoints.openai.serving_models import LoRAModulePath
+from vllm.engine.llm_engine import (LLMEngine, EngineArgs, VllmConfig)
 from vllm.executor.executor_base import ExecutorBase
 
 # Initialize logger
@@ -51,7 +51,7 @@ def _reset_vllm_defaults(self):
         }
         self.vllm_parser.set_defaults(**server_default_args)
 
-        # See https://docs.vllm.ai/en/stable/models/engine_args.html for more args
+        # See https://docs.vllm.ai/en/stable/serving/engine_args.html for more args
         engine_default_args = {
             "model": "/workspace/vllm/weights",
             "cpu_offload_gb": 0,
@@ -121,29 +121,17 @@ def load_lora_adapters(adapters_dir: str) -> Optional[LoRAModulePath]:
 
     return lora_list
 
-def find_max_available_seq_len(engine_config: EngineConfig, max_probe_steps: int) -> int:
+def find_max_available_seq_len(vllm_config: VllmConfig, max_probe_steps: int) -> int:
     """
     Load model and run profiler to find max available seq len.
     """
-    # see https://github.com/vllm-project/vllm/blob/v0.6.3/vllm/engine/llm_engine.py#L335
-    executor_class = LLMEngine._get_executor_cls(engine_config)
-    if engine_config.scheduler_config.enable_chunked_prefill:
+    executor_class = LLMEngine._get_executor_cls(vllm_config)
+    if vllm_config.scheduler_config.enable_chunked_prefill:
         logger.info("Chunked Prefill is enabled, skip probing.")
-        return engine_config.model_config.max_model_len
-    executor = executor_class(
-        model_config=engine_config.model_config,
-        cache_config=engine_config.cache_config,
-        parallel_config=engine_config.parallel_config,
-        scheduler_config=engine_config.scheduler_config,
-        device_config=engine_config.device_config,
-        lora_config=engine_config.lora_config,
-        speculative_config=engine_config.speculative_config,
-        load_config=engine_config.load_config,
-        prompt_adapter_config=engine_config.prompt_adapter_config,
-        observability_config=engine_config.observability_config,
-    )
-
-    res = binary_search_with_limited_steps(engine_config.model_config.max_model_len, max_probe_steps, lambda x: is_context_length_safe(executor, x))
+        return vllm_config.model_config.max_model_len
+    executor = executor_class(vllm_config=vllm_config)
+
+    res = binary_search_with_limited_steps(vllm_config.model_config.max_model_len, max_probe_steps, lambda x: is_context_length_safe(executor, x))
 
     # release memory
     del executor
@@ -188,11 +176,10 @@ def is_context_length_safe(executor: ExecutorBase, context_length: int) -> bool:
     """
     # Profile memory usage with max_num_sequences sequences and the total
     # number of tokens equal to max_num_batched_tokens.
-    # see https://github.com/vllm-project/vllm/blob/v0.6.3/vllm/worker/model_runner.py#L1232
     executor.scheduler_config.max_num_batched_tokens = context_length
     try:
         logger.info(f"Try to determine available gpu blocks for context length {context_length}")
-        # see https://github.com/vllm-project/vllm/blob/v0.6.3/vllm/engine/llm_engine.py#L477
+        # see https://github.com/vllm-project/vllm/blob/v0.7.2/vllm/engine/llm_engine.py#L416
         available_gpu_blocks, _ = executor.determine_num_available_blocks()
     except torch.OutOfMemoryError as e:
         return False    
@@ -223,13 +210,13 @@ def try_set_max_available_seq_len(args: argparse.Namespace):
     # read the model config from hf weights path.
     # vllm will perform different parser for different model architectures
     # and read it into a unified EngineConfig.
-    engine_config = engine_args.create_engine_config()
+    vllm_config = engine_args.create_engine_config()
 
-    max_model_len = engine_config.model_config.max_model_len
+    max_model_len = vllm_config.model_config.max_model_len
     available_seq_len = max_model_len
     logger.info("Try run profiler to find max available seq len")
-    available_seq_len = find_max_available_seq_len(engine_config, max_probe_steps)
-    # see https://github.com/vllm-project/vllm/blob/v0.6.3/vllm/worker/worker.py#L262
+    available_seq_len = find_max_available_seq_len(vllm_config, max_probe_steps)
+    # see https://github.com/vllm-project/vllm/blob/v0.7.2/vllm/worker/worker.py#L539
     if available_seq_len <= 0:
         raise ValueError("No available memory for the cache blocks. "
                         "Try increasing `gpu_memory_utilization` when "