[BugFix][Model] Jamba - Handle aborted requests, Add tests and fix cleanup bug (#6425)

Co-authored-by: Mor Zusman <morz@ai21.com>

tests/models/test_jamba.py

@@ -1,5 +1,6 @@
 import pytest
 
+from tests.models.utils import check_outputs_equal
 from vllm.worker.model_runner import _get_graph_batch_size
 
 MODELS = ["ai21labs/Jamba-tiny-random"]
@@ -34,6 +35,34 @@ def test_models(
             f"Test{i}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}")
 
 
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [20])
+def test_batching(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    # To pass the small model tests, we need full precision.
+    for_loop_outputs = []
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        for prompt in example_prompts:
+            for_loop_outputs.append(
+                vllm_model.generate_greedy([prompt], max_tokens)[0])
+
+        batched_outputs = vllm_model.generate_greedy(example_prompts,
+                                                     max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=for_loop_outputs,
+        outputs_1_lst=batched_outputs,
+        name_0="for_loop_vllm",
+        name_1="batched_vllm",
+    )
+
+
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["bfloat16"])
 @pytest.mark.parametrize("max_tokens", [20])
@@ -60,6 +89,60 @@ def test_mamba_cache_cg_padding(
             "Could be related to mamba cache not padded correctly")
 
 
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [20])
+def test_models_preemption_recompute(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    # Tests that outputs are identical with and w/o preemtions (recompute)
+    assert dtype == "float"
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_model.model.llm_engine.scheduler[
+            0].ENABLE_ARTIFICIAL_PREEMPT = True
+        preempt_vllm_outputs = vllm_model.generate_greedy(
+            example_prompts, max_tokens)
+
+        vllm_model.model.llm_engine.scheduler[
+            0].ENABLE_ARTIFICIAL_PREEMPT = False
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=preempt_vllm_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="vllm_preepmtions",
+        name_1="vllm",
+    )
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+def test_fail_upon_inc_requests_and_finished_requests_lt_available_blocks(
+    vllm_runner,
+    model: str,
+    dtype: str,
+    example_prompts,
+) -> None:
+    # This test is for verifying that the Jamba inner state management doesn't
+    # collapse in case where the number of incoming requests and
+    # finished_requests_ids is larger than the maximum mamba block capacity.
+    # This could generally happen due to the fact that Jamba does support
+    # statelessness mechanism where it can cleanup new incoming requests in
+    # a single step.
+    try:
+        with vllm_runner(model, dtype=dtype, max_num_seqs=10) as vllm_model:
+            vllm_model.generate_greedy([example_prompts[0]] * 100, 10)
+    except ValueError:
+        pytest.fail("Jamba inner state wasn't cleaned up properly between"
+                    "steps finished requests registered unnecessarily ")
+
+
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["float"])
 def test_state_cleanup(

vllm/core/scheduler.py

@@ -374,6 +374,7 @@ class Scheduler:
             for aborted_group in aborted_groups:
                 # Remove the sequence group from the state queue.
                 state_queue.remove(aborted_group)
+                self._finished_requests_ids.append(aborted_group.request_id)
                 for seq in aborted_group.get_seqs():
                     if seq.is_finished():
                         continue

vllm/model_executor/model_loader/loader.py

@@ -32,7 +32,8 @@ from vllm.model_executor.model_loader.weight_utils import (
     filter_duplicate_safetensors_files, filter_files_not_needed_for_inference,
     get_quant_config, initialize_dummy_weights, np_cache_weights_iterator,
     pt_weights_iterator, safetensors_weights_iterator)
-from vllm.model_executor.models.interfaces import (supports_lora,
+from vllm.model_executor.models.interfaces import (has_inner_state,
+                                                   supports_lora,
                                                    supports_vision)
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
@@ -69,7 +70,7 @@ def _get_model_initialization_kwargs(
         model_class: Type[nn.Module],
         lora_config: Optional[LoRAConfig],
         multimodal_config: Optional[MultiModalConfig],
-) -> Dict[str, Any]:
+        scheduler_config: Optional[SchedulerConfig] = None) -> Dict[str, Any]:
     """Get extra kwargs for model initialization."""
     extra_kwargs: Dict[str, Any] = {}
 
@@ -90,13 +91,19 @@ def _get_model_initialization_kwargs(
 
         extra_kwargs["multimodal_config"] = multimodal_config
 
+    if has_inner_state(model_class) and scheduler_config:
+        extra_kwargs["scheduler_config"] = scheduler_config
+
     return extra_kwargs
 
 
-def _initialize_model(model_config: ModelConfig, load_config: LoadConfig,
+def _initialize_model(
+        model_config: ModelConfig,
+        load_config: LoadConfig,
         lora_config: Optional[LoRAConfig],
         multimodal_config: Optional[MultiModalConfig],
-                      cache_config: CacheConfig) -> nn.Module:
+        cache_config: CacheConfig,
+        scheduler_config: Optional[SchedulerConfig] = None) -> nn.Module:
     """Initialize a model with the given configurations."""
     model_class = get_model_architecture(model_config)[0]
     quant_config = _get_quantization_config(model_config, load_config)
@@ -105,7 +112,8 @@ def _initialize_model(model_config: ModelConfig, load_config: LoadConfig,
                        cache_config=cache_config,
                        quant_config=quant_config,
                        **_get_model_initialization_kwargs(
-                           model_class, lora_config, multimodal_config))
+                           model_class, lora_config, multimodal_config,
+                           scheduler_config))
 
 
 class BaseModelLoader(ABC):
@@ -266,7 +274,7 @@ class DefaultModelLoader(BaseModelLoader):
             with torch.device(device_config.device):
                 model = _initialize_model(model_config, self.load_config,
                                           lora_config, multimodal_config,
-                                          cache_config)
+                                          cache_config, scheduler_config)
             model.load_weights(
                 self._get_weights_iterator(model_config.model,
                                            model_config.revision,
@@ -302,7 +310,7 @@ class DummyModelLoader(BaseModelLoader):
             with torch.device(device_config.device):
                 model = _initialize_model(model_config, self.load_config,
                                           lora_config, multimodal_config,
-                                          cache_config)
+                                          cache_config, scheduler_config)
             # NOTE(woosuk): For accurate performance evaluation, we assign
             # random values to the weights.
             initialize_dummy_weights(model)

vllm/model_executor/models/interfaces.py

@@ -3,7 +3,7 @@ from typing import (ClassVar, Dict, List, Literal, Optional, Protocol, Type,
 
 from typing_extensions import TypeGuard
 
-from vllm.config import LoRAConfig, MultiModalConfig
+from vllm.config import LoRAConfig, MultiModalConfig, SchedulerConfig
 from vllm.logger import init_logger
 
 logger = init_logger(__name__)
@@ -142,3 +142,49 @@ def _supports_lora(
         return isinstance(model, _SupportsLoRAType)
 
     return isinstance(model, SupportsLoRA)
+
+
+@runtime_checkable
+class HasInnerState(Protocol):
+    """The interface required for all models that has inner state."""
+
+    has_inner_state: ClassVar[Literal[True]] = True
+    """
+        A flag that indicates this model has inner state.
+        Models that has inner state usually need access to the scheduler_config
+        for max_num_seqs ,etc... (Currently only used by Jamba)
+    """
+
+    def __init__(self,
+                 *,
+                 scheduler_config: Optional[SchedulerConfig] = None) -> None:
+        ...
+
+
+@runtime_checkable
+class _HasInnerStateType(Protocol):
+    has_inner_state: ClassVar[Literal[True]]
+
+    def __init__(self,
+                 *,
+                 scheduler_config: Optional[SchedulerConfig] = None) -> None:
+        ...
+
+
+@overload
+def has_inner_state(model: object) -> TypeGuard[HasInnerState]:
+    ...
+
+
+@overload
+def has_inner_state(model: Type[object]) -> TypeGuard[Type[HasInnerState]]:
+    ...
+
+
+def has_inner_state(
+    model: Union[Type[object], object]
+) -> Union[TypeGuard[Type[HasInnerState]], TypeGuard[HasInnerState]]:
+    if isinstance(model, type):
+        return isinstance(model, _HasInnerStateType)
+
+    return isinstance(model, HasInnerState)

vllm/model_executor/models/jamba.py

@@ -13,7 +13,7 @@ from transformers import JambaConfig
 
 from vllm.attention.backends.abstract import AttentionMetadata
 from vllm.attention.layer import Attention
-from vllm.config import CacheConfig, LoRAConfig
+from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig
 from vllm.distributed import (get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size,
                               tensor_model_parallel_all_reduce)
@@ -32,10 +32,12 @@ from vllm.model_executor.layers.sampler import Sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.models.interfaces import HasInnerState
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.sequence import IntermediateTensors, SamplerOutput
-from vllm.worker.model_runner import _BATCH_SIZES_TO_CAPTURE
+from vllm.worker.model_runner import (_BATCH_SIZES_TO_CAPTURE,
+                                      _get_graph_batch_size)
 
 KVCache = Tuple[torch.Tensor, torch.Tensor]
 
@@ -612,7 +614,7 @@ class JambaModel(nn.Module):
         return hidden_states
 
 
-class JambaForCausalLM(nn.Module):
+class JambaForCausalLM(nn.Module, HasInnerState):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -640,9 +642,11 @@ class JambaForCausalLM(nn.Module):
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         lora_config: Optional[LoRAConfig] = None,
+        scheduler_config: Optional[SchedulerConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
+        self.scheduler_config = scheduler_config
         self.model = JambaModel(config,
                                 cache_config=cache_config,
                                 quant_config=quant_config,
@@ -689,6 +693,8 @@ class JambaForCausalLM(nn.Module):
                 for key in ["request_ids_to_seq_ids", "finished_requests_ids"])
 
             request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
+            finished_requests_ids = kwargs["finished_requests_ids"]
+            self._release_mamba_cache(finished_requests_ids)
             batch_size = input_ids.shape[0]
             if attn_metadata.prefill_metadata:
                 batch_size = len(request_ids_to_seq_ids)
@@ -696,9 +702,8 @@ class JambaForCausalLM(nn.Module):
                 current_seqlen_agnostic_cache,
                 indices,
             ) = self._prepare_current_run_mamba_cache(request_ids_to_seq_ids,
-                                                      batch_size)
+                                                      batch_size,
-            finished_requests_ids = kwargs["finished_requests_ids"]
+                                                      finished_requests_ids)
-            self._release_mamba_cache(finished_requests_ids)
         else:
             # CUDA graph capturing runs
             current_seqlen_agnostic_cache, indices = (
@@ -760,10 +765,15 @@ class JambaForCausalLM(nn.Module):
         return indices_for_current_run
 
     def _prepare_current_run_mamba_cache(
-        self, request_ids_to_seq_ids: Dict[str, list[int]], batch_size: int
+        self, request_ids_to_seq_ids: Dict[str, list[int]], batch_size: int,
+        finished_requests_ids: List[str]
     ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], List[int]]:
         indices_for_current_run = []
         for request_id, seqs_id in request_ids_to_seq_ids.items():
+            if request_id in finished_requests_ids:
+                # Do not allocate cache for requests that run
+                # and finish right after
+                continue
             indices_for_current_run += self._assign_seq_id_to_mamba_cache(
                 request_id, seqs_id)
         ## Pad the batch in case of running batch that was not captured via CG
@@ -787,16 +797,17 @@ class JambaForCausalLM(nn.Module):
         assert all(
             key in kwargs
             for key in ["request_ids_to_seq_ids", "finished_requests_ids"])
+        finished_requests_ids = kwargs["finished_requests_ids"]
+        self._release_mamba_cache(finished_requests_ids)
         request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
         cg_batch_size = input_buffers['input_ids'].shape[0]
         (
             current_mamba_cache,
             indices,
         ) = self._prepare_current_run_mamba_cache(request_ids_to_seq_ids,
-                                                  cg_batch_size)
+                                                  cg_batch_size,
+                                                  finished_requests_ids)
         self.current_indices = indices
-        finished_requests_ids = kwargs["finished_requests_ids"]
-        self._release_mamba_cache(finished_requests_ids)
 
         for input_buffer, current_cache_buffer in zip(
                 input_buffers["seqlen_agnostic_capture_inputs"],
@@ -860,9 +871,12 @@ class JambaForCausalLM(nn.Module):
         layers_type = self.config.layers_block_type
         mamba_layers = sum(
             [layer_type == "mamba" for layer_type in layers_type])
-        max_batch_size = _BATCH_SIZES_TO_CAPTURE[-1] + 10
+        max_batch_size = (_get_graph_batch_size(
+            self.scheduler_config.max_num_seqs) if self.scheduler_config else
+                          max(_BATCH_SIZES_TO_CAPTURE)) + 10
         conv_state_shape, temporal_state_shape = self._get_mamba_cache_shape()
         assert conv_state_shape is not None and temporal_state_shape is not None
+
         for buffername in ["mamba_cache", "mamba_gc_cache_buffer"]:
             buffer = (torch.empty(size=(mamba_layers, max_batch_size) +
                                   conv_state_shape,