support pp virtual engine

Signed-off-by: Chen Zhang <[email protected]>

Author: heheda12345

vllm/attention/layer.py

@@ -123,7 +123,11 @@ def __init__(
         self.attn_type = attn_type
         # use a placeholder kv cache tensor during init, which will be replaced
         # by bind_kv_cache
-        self.kv_cache = torch.tensor([])
+        # this variable will not be accessed if use_direct_call is True
+        self.kv_cache = [
+            torch.tensor([]) for _ in range(get_current_vllm_config(
+            ).parallel_config.pipeline_parallel_size)
+        ]
 
     def forward(
         self,
@@ -238,7 +242,8 @@ def unified_attention(
     forward_context: ForwardContext = get_forward_context()
     attn_metadata = forward_context.attn_metadata
     self = forward_context.attn_layers[layer_name]
-    return self.impl.forward(query, key, value, self.kv_cache, attn_metadata,
+    kv_cache = self.kv_cache[forward_context.virtual_engine]
+    return self.impl.forward(query, key, value, kv_cache, attn_metadata,
                              self._k_scale, self._v_scale)
 
 
@@ -270,10 +275,11 @@ def unified_attention_with_output(
     forward_context: ForwardContext = get_forward_context()
     attn_metadata = forward_context.attn_metadata
     self = forward_context.attn_layers[layer_name]
+    kv_cache = self.kv_cache[forward_context.virtual_engine]
     self.impl.forward(query,
                       key,
                       value,
-                      self.kv_cache,
+                      kv_cache,
                       attn_metadata,
                       self._k_scale,
                       self._v_scale,

vllm/forward_context.py

@@ -28,6 +28,8 @@ class ForwardContext:
     attn_layers: Dict[str, Any]
     # TODO: extend to support per-layer dynamic forward context
     attn_metadata: "AttentionMetadata"  # set dynamically for each forward pass
+    # TODO: remove after making all virtual_engines share the same kv cache
+    virtual_engine: int  # set dynamically for each forward pass
 
 
 _forward_context: Optional[ForwardContext] = None
@@ -42,7 +44,9 @@ def get_forward_context() -> ForwardContext:
 
 
 @contextmanager
-def set_forward_context(attn_metadata: Any, vllm_config: VllmConfig):
+def set_forward_context(attn_metadata: Any,
+                        vllm_config: VllmConfig,
+                        virtual_engine: int = 0):
     """A context manager that stores the current forward context,
     can be attention metadata, etc.
     Here we can inject common logic for every model forward pass.
@@ -55,6 +59,7 @@ def set_forward_context(attn_metadata: Any, vllm_config: VllmConfig):
     prev_context = _forward_context
     _forward_context = ForwardContext(
         attn_layers=vllm_config.compilation_config.static_forward_context,
+        virtual_engine=virtual_engine,
         attn_metadata=attn_metadata)
     try:
         yield

vllm/utils.py

@@ -1949,16 +1949,20 @@ def get_mp_context():
     return multiprocessing.get_context(mp_method)
 
 
-def bind_kv_cache(ctx: Dict[str, Any], kv_cache: List[torch.Tensor]) -> None:
+def bind_kv_cache(
+        ctx: Dict[str, Any],
+        kv_cache: List[List[torch.Tensor]],  # [virtual_engine][layer_index]
+) -> None:
     # Bind the kv_cache tensor to Attention modules, similar to
-    # ctx[layer_name].kv_cache = kv_cache[extract_layer_index(layer_name)]
+    # ctx[layer_name].kv_cache[ve]=kv_cache[ve][extract_layer_index(layer_name)]
     # Special things handled here:
     # 1. Some models have non-attention layers, e.g., Jamba
     # 2. Pipeline parallelism, each rank only has a subset of layers
     # 3. Encoder attention has no kv cache
     # 4. Encoder-decoder models, encoder-decoder attention and decoder-only
     #    attention of the same layer (e.g., bart's decoder.layers.1.self_attn
-    #    and decoder.layers.1.encoder_attn is mapped to the same kv cache tensor
+    #    and decoder.layers.1.encoder_attn) is mapped to the same kv cache
+    #    tensor
     from vllm.attention import AttentionType
     from vllm.model_executor.models.utils import extract_layer_index
     layer_need_kv_cache = [
@@ -1974,4 +1978,7 @@ def bind_kv_cache(ctx: Dict[str, Any], kv_cache: List[torch.Tensor]) -> None:
         kv_cache_idx = layer_index_sorted.index(
             extract_layer_index(layer_name))
         forward_ctx = ctx[layer_name]
-        forward_ctx.kv_cache = kv_cache[kv_cache_idx]
+        assert len(forward_ctx.kv_cache) == len(kv_cache)
+        for ve, ve_kv_cache in enumerate(kv_cache):
+            assert forward_ctx.kv_cache[ve].numel() == 0
+            forward_ctx.kv_cache[ve] = ve_kv_cache[kv_cache_idx]

vllm/v1/worker/gpu_model_runner.py

@@ -863,4 +863,4 @@ def initialize_kv_cache(self, num_blocks: int) -> None:
                             device=self.device))
         bind_kv_cache(
             self.vllm_config.compilation_config.static_forward_context,
-            self.kv_caches)
+            [self.kv_caches])

vllm/worker/cache_engine.py

@@ -4,11 +4,10 @@
 import torch
 
 from vllm.attention import get_attn_backend
-from vllm.config import (CacheConfig, CompilationConfig, DeviceConfig,
-                         ModelConfig, ParallelConfig)
+from vllm.config import CacheConfig, DeviceConfig, ModelConfig, ParallelConfig
 from vllm.logger import init_logger
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, LayerBlockType,
-                        bind_kv_cache, get_dtype_size, is_pin_memory_available)
+                        get_dtype_size, is_pin_memory_available)
 
 logger = init_logger(__name__)
 
@@ -21,14 +20,9 @@ class CacheEngine:
     as swapping and copying.
     """
 
-    def __init__(
-        self,
-        cache_config: CacheConfig,
-        model_config: ModelConfig,
-        parallel_config: ParallelConfig,
-        device_config: DeviceConfig,
-        compilation_config: CompilationConfig,
-    ) -> None:
+    def __init__(self, cache_config: CacheConfig, model_config: ModelConfig,
+                 parallel_config: ParallelConfig,
+                 device_config: DeviceConfig) -> None:
         self.cache_config = cache_config
         self.model_config = model_config
         self.parallel_config = parallel_config
@@ -64,8 +58,6 @@ def __init__(
         self.gpu_cache = self._allocate_kv_cache(
             self.num_gpu_blocks, self.device_config.device_type)
         self.cpu_cache = self._allocate_kv_cache(self.num_cpu_blocks, "cpu")
-        bind_kv_cache(compilation_config.static_forward_context,
-                      self.gpu_cache)
 
     def _allocate_kv_cache(
         self,

vllm/worker/cpu_enc_dec_model_runner.py

@@ -305,7 +305,8 @@ def execute_model(
             intermediate_tensors,
         }
 
-        with set_forward_context(model_input.attn_metadata, self.vllm_config):
+        with set_forward_context(model_input.attn_metadata, self.vllm_config,
+                                 model_input.virtual_engine):
             hidden_states = model_executable(**execute_model_kwargs)
 
         # Compute the logits.

vllm/worker/cpu_model_runner.py

@@ -526,7 +526,8 @@ def execute_model(
             execute_model_kwargs.update(
                 {"previous_hidden_states": previous_hidden_states})
 
-        with set_forward_context(model_input.attn_metadata, self.vllm_config):
+        with set_forward_context(model_input.attn_metadata, self.vllm_config,
+                                 model_input.virtual_engine):
             hidden_states = model_executable(
                 input_ids=model_input.input_tokens,
                 positions=model_input.input_positions,

vllm/worker/cpu_pooling_model_runner.py

@@ -69,7 +69,8 @@ def execute_model(
             intermediate_tensors,
         }
 
-        with set_forward_context(model_input.attn_metadata, self.vllm_config):
+        with set_forward_context(model_input.attn_metadata, self.vllm_config,
+                                 model_input.virtual_engine):
             hidden_states = model_executable(**execute_model_kwargs)
 
         # Only perform pooling in the driver worker.

vllm/worker/cpu_worker.py

@@ -6,8 +6,8 @@
 
 import vllm.envs as envs
 from vllm.attention import get_attn_backend
-from vllm.config import (CacheConfig, CompilationConfig, DeviceConfig,
-                         ModelConfig, ParallelConfig, VllmConfig)
+from vllm.config import (CacheConfig, DeviceConfig, ModelConfig,
+                         ParallelConfig, VllmConfig)
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
 from vllm.logger import init_logger
@@ -33,8 +33,8 @@ class CPUCacheEngine:
     """
 
     def __init__(self, cache_config: CacheConfig, model_config: ModelConfig,
-                 parallel_config: ParallelConfig, device_config: DeviceConfig,
-                 compilation_config: CompilationConfig) -> None:
+                 parallel_config: ParallelConfig,
+                 device_config: DeviceConfig) -> None:
         assert device_config.device_type == "cpu"
         self.cache_config = cache_config
         self.model_config = model_config
@@ -66,8 +66,6 @@ def __init__(self, cache_config: CacheConfig, model_config: ModelConfig,
 
         # Initialize the cache.
         self.cpu_cache = self._allocate_kv_cache(self.num_cpu_blocks)
-        bind_kv_cache(compilation_config.static_forward_context,
-                      self.cpu_cache)
 
     def _allocate_kv_cache(
         self,
@@ -292,13 +290,15 @@ def _init_cache_engine(self) -> None:
                 self.model_config,
                 self.parallel_config,
                 self.device_config,
-                self.compilation_config,
             ) for _ in range(self.parallel_config.pipeline_parallel_size)
         ]
         self.cpu_cache = [
             self.cache_engine[ve].cpu_cache
             for ve in range(self.parallel_config.pipeline_parallel_size)
         ]
+        for ve in range(self.parallel_config.pipeline_parallel_size):
+            bind_kv_cache(self.compilation_config.static_forward_context,
+                          self.cpu_cache[ve], ve)
         self.model_runner.block_size = self.cache_engine[0].block_size
 
         assert all(

vllm/worker/enc_dec_model_runner.py

@@ -175,7 +175,8 @@ def execute_model(
         } if self.has_inner_state else {}
 
         multi_modal_kwargs = model_input.multi_modal_kwargs or {}
-        with set_forward_context(model_input.attn_metadata, self.vllm_config):
+        with set_forward_context(model_input.attn_metadata, self.vllm_config,
+                                 model_input.virtual_engine):
             hidden_or_intermediate_states = model_executable(
                 input_ids=model_input.input_tokens,
                 positions=model_input.input_positions,