[Model] Add user-configurable task for models that support both gener…

…ation and embedding (#9424)

docs/source/models/supported_models.rst

@@ -294,6 +294,10 @@ Text Embedding
     - 
     - ✅︎
 
+.. important::
+  Some model architectures support both generation and embedding tasks.
+  In this case, you have to pass :code:`--task embedding` to run the model in embedding mode.
+
 Reward Modeling
 ---------------
 
@@ -482,6 +486,10 @@ Multimodal Embedding
     - 🚧
     - ✅︎
 
+.. important::
+  Some model architectures support both generation and embedding tasks.
+  In this case, you have to pass :code:`--task embedding` to run the model in embedding mode.
+
 ----
 
 If your model uses one of the above model architectures, you can seamlessly run your model with vLLM.

docs/source/models/vlm.rst

@@ -181,8 +181,8 @@ Below is an example on how to launch the same ``microsoft/Phi-3.5-vision-instruc
 
 .. code-block:: bash
 
-    vllm serve microsoft/Phi-3.5-vision-instruct --max-model-len 4096 \
-      --trust-remote-code --limit-mm-per-prompt image=2
+    vllm serve microsoft/Phi-3.5-vision-instruct --task generate \
+      --trust-remote-code --max-model-len 4096 --limit-mm-per-prompt image=2
 
 .. important::
     Since OpenAI Vision API is based on `Chat Completions <https://platform.openai.com/docs/api-reference/chat>`_ API,

examples/offline_inference_vision_language_embedding.py

@@ -7,6 +7,7 @@
 # Create an LLM.
 llm = LLM(
     model="TIGER-Lab/VLM2Vec-Full",
+    task="embedding",
     trust_remote_code=True,
     max_model_len=4096,
     max_num_seqs=2,

examples/openai_api_client_for_multimodal.py

@@ -7,8 +7,8 @@
 vllm serve llava-hf/llava-1.5-7b-hf --chat-template template_llava.jinja
 
 (multi-image inference with Phi-3.5-vision-instruct)
-vllm serve microsoft/Phi-3.5-vision-instruct --max-model-len 4096 \
-    --trust-remote-code --limit-mm-per-prompt image=2
+vllm serve microsoft/Phi-3.5-vision-instruct --task generate \
+    --trust-remote-code --max-model-len 4096 --limit-mm-per-prompt image=2
 
 (audio inference with Ultravox)
 vllm serve fixie-ai/ultravox-v0_3 --max-model-len 4096

tests/conftest.py

@@ -25,7 +25,7 @@
 from vllm import LLM, SamplingParams
 from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
-from vllm.config import TokenizerPoolConfig
+from vllm.config import TaskOption, TokenizerPoolConfig
 from vllm.connections import global_http_connection
 from vllm.distributed import (destroy_distributed_environment,
                               destroy_model_parallel,
@@ -619,6 +619,7 @@ class VllmRunner:
     def __init__(
         self,
         model_name: str,
+        task: TaskOption = "auto",
         tokenizer_name: Optional[str] = None,
         # Use smaller max model length, otherwise bigger model cannot run due
         # to kv cache size limit.
@@ -634,6 +635,7 @@ def __init__(
     ) -> None:
         self.model = LLM(
             model=model_name,
+            task=task,
             tokenizer=tokenizer_name,
             trust_remote_code=True,
             dtype=dtype,

tests/core/test_chunked_prefill_scheduler.py

@@ -33,7 +33,8 @@ def test_simple():
     num_seq_group = 4
     max_model_len = 16
     max_num_batched_tokens = 64
-    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+    scheduler_config = SchedulerConfig("generate",
+                                       max_num_batched_tokens,
                                        num_seq_group,
                                        max_model_len,
                                        enable_chunked_prefill=True)
@@ -78,6 +79,7 @@ def test_chunk():
     max_model_len = 80
     max_num_batched_tokens = 64
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -126,6 +128,7 @@ def test_complex():
     max_model_len = 80
     max_num_batched_tokens = 64
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -196,6 +199,7 @@ def test_maximal_decoding():
     max_model_len = 8
     max_num_batched_tokens = 2
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -289,6 +293,7 @@ def test_prompt_limit():
     max_model_len = 64
     max_num_batched_tokens = 32
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -321,7 +326,8 @@ def test_prompt_limit_exceed():
     max_seqs = 64
     max_model_len = 32
     max_num_batched_tokens = 64
-    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+    scheduler_config = SchedulerConfig("generate",
+                                       max_num_batched_tokens,
                                        max_seqs,
                                        max_model_len,
                                        enable_chunked_prefill=True)
@@ -348,6 +354,7 @@ def test_swap():
     max_model_len = 200
     max_num_batched_tokens = 30
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -404,6 +411,7 @@ def test_running_prefill_prioritized_over_swap():
     max_model_len = 200
     max_num_batched_tokens = 30
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -498,6 +506,7 @@ def test_chunked_prefill_preempt():
     max_model_len = 200
     max_num_batched_tokens = 30
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -563,6 +572,7 @@ def test_chunked_prefill_max_seqs():
     max_model_len = 80
     max_num_batched_tokens = 64
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,
@@ -617,6 +627,7 @@ def test_perfix_caching():
     max_model_len = 80
     max_num_batched_tokens = 64
     scheduler_config = SchedulerConfig(
+        "generate",
         max_num_batched_tokens,
         max_seqs,
         max_model_len,

tests/core/test_scheduler.py

@@ -20,9 +20,10 @@
 def test_scheduler_add_seq_group():
     block_size = 4
     scheduler_config = SchedulerConfig(
-        100,
-        64,
-        1,
+        "generate",
+        max_num_batched_tokens=100,
+        max_num_seqs=64,
+        max_model_len=1,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, cache_dtype="auto")
     cache_config.num_cpu_blocks = 4
@@ -42,9 +43,10 @@ def test_scheduler_add_seq_group():
 def test_scheduler_abort_seq_group():
     block_size = 4
     scheduler_config = SchedulerConfig(
-        100,
-        64,
-        1,
+        "generate",
+        max_num_batched_tokens=100,
+        max_num_seqs=64,
+        max_model_len=1,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = 4
@@ -70,9 +72,10 @@ def test_scheduler_schedule_simple():
     num_seq_group = 4
     max_model_len = 16
     scheduler_config = SchedulerConfig(
-        64,
-        num_seq_group,
-        max_model_len,
+        "generate",
+        max_num_batched_tokens=64,
+        max_num_seqs=num_seq_group,
+        max_model_len=max_model_len,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = 8
@@ -114,9 +117,10 @@ def test_scheduler_prefill_prioritized():
     max_model_len = 30
     max_batched_num_tokens = 30
     scheduler_config = SchedulerConfig(
-        max_batched_num_tokens,
-        2,
-        max_model_len,
+        "generate",
+        max_num_batched_tokens=max_batched_num_tokens,
+        max_num_seqs=2,
+        max_model_len=max_model_len,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = 16
@@ -145,9 +149,10 @@ def test_scheduler_schedule_preempt_abort():
     block_size = 4
     max_model_len = 16
     scheduler_config = SchedulerConfig(
-        64,
-        2,
-        max_model_len,
+        "generate",
+        max_num_batched_tokens=64,
+        max_num_seqs=2,
+        max_model_len=max_model_len,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = 2
@@ -204,9 +209,10 @@ def test_scheduler_max_seqs():
     max_seq_group = 2
     max_model_len = 16
     scheduler_config = SchedulerConfig(
-        64,
-        max_seq_group,
-        max_model_len,
+        "generate",
+        max_num_batched_tokens=64,
+        max_num_seqs=max_seq_group,
+        max_model_len=max_model_len,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = 8
@@ -248,9 +254,10 @@ def test_scheduler_max_seqs():
 def test_scheduler_delay_factor():
     block_size = 4
     scheduler_config = SchedulerConfig(
-        100,
-        64,
-        16,
+        "generate",
+        max_num_batched_tokens=100,
+        max_num_seqs=64,
+        max_model_len=16,
         delay_factor=0.5,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
@@ -350,9 +357,10 @@ def initialize_scheduler(
 ):
     block_size = block_size
     scheduler_config = SchedulerConfig(
-        max_token_budget,
-        max_num_seqs,
-        max_model_len,
+        "generate",
+        max_num_batched_tokens=max_token_budget,
+        max_num_seqs=max_num_seqs,
+        max_model_len=max_model_len,
     )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = num_cpu_blocks

tests/core/test_scheduler_encoder_decoder.py

@@ -36,7 +36,12 @@ def test_scheduler_schedule_simple_encoder_decoder():
     block_size = 4
     num_seq_group = 4
     max_model_len = 16
-    scheduler_config = SchedulerConfig(64, num_seq_group, max_model_len)
+    scheduler_config = SchedulerConfig(
+        task="generate",
+        max_num_batched_tokens=64,
+        max_num_seqs=num_seq_group,
+        max_model_len=max_model_len,
+    )
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
     cache_config.num_cpu_blocks = 16  # enc and dec prompts per seq_group
     cache_config.num_gpu_blocks = 16  # enc and dec prompts per seq_group