[V1] Support cross-layer KV sharing (#18212)

Signed-off-by: Yong Hoon Shin <yhshin@meta.com>

vllm/v1/attention/backends/flashinfer.py

@@ -507,6 +507,7 @@ class FlashInferImpl(AttentionImpl):
         blocksparse_params: Optional[dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: Optional[int] = None,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -521,6 +522,7 @@ class FlashInferImpl(AttentionImpl):
             self.sliding_window = (sliding_window - 1, 0)
         self.kv_cache_dtype = kv_cache_dtype
         self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
 
         assert self.num_heads % self.num_kv_heads == 0
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
@@ -568,11 +570,15 @@ class FlashInferImpl(AttentionImpl):
         # performance to make sure it does not introduce any overhead.
 
         num_actual_tokens = attn_metadata.num_actual_tokens
+
+        if self.kv_sharing_target_layer_name is None:
             # Reshape the input keys and values and store them in the cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
             # NOTE(woosuk): Here, key and value are padded while slot_mapping is
-        # not padded. However, we don't need to do key[:num_actual_tokens] and
-        # value[:num_actual_tokens] because the reshape_and_cache_flash op uses
-        # the slot_mapping's shape to determine the number of actual tokens.
+            # not padded. However, we don't need to do key[:num_actual_tokens]
+            # and value[:num_actual_tokens] because the reshape_and_cache_flash
+            # op uses the slot_mapping's shape to determine the number of
+            # actual tokens.
             torch.ops._C_cache_ops.reshape_and_cache_flash(
                 key,
                 value,

vllm/v1/attention/backends/mla/common.py

@@ -586,6 +586,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
         blocksparse_params: Optional[dict[str, Any]],
         logits_soft_cap: Optional[float],
         attn_type: str,
+        kv_sharing_target_layer_name: Optional[str],
         # MLA Specific Arguments
         q_lora_rank: Optional[int],
         kv_lora_rank: int,
@@ -595,6 +596,9 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
         v_head_dim: int,
         kv_b_proj: ColumnParallelLinear,
     ) -> None:
+        if kv_sharing_target_layer_name is not None:
+            raise NotImplementedError("KV sharing is not supported for MLA")
+
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/v1/attention/backends/mla/flashmla.py

@@ -93,12 +93,13 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
             blocksparse_params: Optional[dict[str, Any]],
             logits_soft_cap: Optional[float],
             attn_type: str,
+            kv_sharing_target_layer_name: Optional[str],
             # MLA Specific Arguments
             **mla_args) -> None:
         super().__init__(num_heads, head_size, scale, num_kv_heads,
                          alibi_slopes, sliding_window, kv_cache_dtype,
                          blocksparse_params, logits_soft_cap, attn_type,
-                         **mla_args)
+                         kv_sharing_target_layer_name, **mla_args)
 
         assert is_flashmla_supported(), \
             "FlashMLA is not supported on this device"

vllm/v1/attention/backends/mla/rocm_aiter_mla.py

@@ -139,12 +139,13 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
             blocksparse_params: Optional[dict[str, Any]],
             logits_soft_cap: Optional[float],
             attn_type: str,
+            kv_sharing_target_layer_name: Optional[str],
             # MLA Specific Arguments
             **mla_args) -> None:
         super().__init__(num_heads, head_size, scale, num_kv_heads,
                          alibi_slopes, sliding_window, kv_cache_dtype,
                          blocksparse_params, logits_soft_cap, attn_type,
-                         **mla_args)
+                         kv_sharing_target_layer_name, **mla_args)
         assert (num_heads == 16 or num_heads == 128), (
             f"Aiter MLA only supports 16 or 128 number of heads.\n"
             f"Provided {num_heads} number of heads.\n"

vllm/v1/attention/backends/mla/triton_mla.py

@@ -41,12 +41,13 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
             blocksparse_params: Optional[dict[str, Any]],
             logits_soft_cap: Optional[float],
             attn_type: str,
+            kv_sharing_target_layer_name: Optional[str],
             # MLA Specific Arguments
             **mla_args) -> None:
         super().__init__(num_heads, head_size, scale, num_kv_heads,
                          alibi_slopes, sliding_window, kv_cache_dtype,
                          blocksparse_params, logits_soft_cap, attn_type,
-                         **mla_args)
+                         kv_sharing_target_layer_name, **mla_args)
 
         unsupported_features = [
             alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap

vllm/v1/attention/backends/pallas.py

@@ -113,6 +113,7 @@ class PallasAttentionBackendImpl(AttentionImpl):
         blocksparse_params: Optional[dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        kv_sharing_target_layer_name: Optional[int] = None,
         use_irope: bool = False,
     ) -> None:
         if use_irope:
@@ -128,6 +129,7 @@ class PallasAttentionBackendImpl(AttentionImpl):
         self.num_kv_heads = num_kv_heads
         self.sliding_window = sliding_window
         self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
 
         assert self.num_heads % self.num_kv_heads == 0
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
@@ -181,7 +183,9 @@ class PallasAttentionBackendImpl(AttentionImpl):
         num_tokens, hidden_size = query.shape
         query = query.view(num_tokens, self.num_heads, self.head_size)
 
-        if kv_cache.numel() > 0:
+        if self.kv_sharing_target_layer_name is None and kv_cache.numel() > 0:
+            # Write input keys and values to the KV cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
             slot_mapping = attn_metadata.slot_mapping
             write_to_kv_cache(key, value, kv_cache, slot_mapping)
 

vllm/v1/attention/backends/triton_attn.py

@@ -88,6 +88,7 @@ class TritonAttentionImpl(AttentionImpl):
         blocksparse_params: Optional[dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
+        kv_sharing_target_layer_name: Optional[int] = None,
         use_irope: bool = False,
     ) -> None:
         if blocksparse_params is not None:
@@ -109,6 +110,7 @@ class TritonAttentionImpl(AttentionImpl):
             # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
             logits_soft_cap = 0
         self.logits_soft_cap = logits_soft_cap
+        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
 
         self.use_irope = use_irope
 
@@ -178,8 +180,13 @@ class TritonAttentionImpl(AttentionImpl):
         if use_prefill_decode_attn:
             key_cache, value_cache = PagedAttention.split_kv_cache(
                 kv_cache, self.num_kv_heads, self.head_size)
+        else:
+            key_cache, value_cache = kv_cache.unbind(0)
 
+        if self.kv_sharing_target_layer_name is None:
             # Reshape the input keys and values and store them in the cache.
+            # Skip this if sharing KV cache with an earlier attention layer.
+            if use_prefill_decode_attn:
                 PagedAttention.write_to_paged_cache(
                     key,
                     value,
@@ -190,9 +197,7 @@ class TritonAttentionImpl(AttentionImpl):
                     layer._k_scale,
                     layer._v_scale,
                 )
-
             else:
-            key_cache, value_cache = kv_cache.unbind(0)
                 torch.ops._C_cache_ops.reshape_and_cache_flash(
                     key,
                     value,

vllm/v1/attention/backends/utils.py

@@ -17,3 +17,36 @@ class CommonAttentionMetadata:
     seq_lens: torch.Tensor
     """(batch_size,), the length of each request including both computed tokens
     and newly scheduled tokens"""
+
+
+def validate_kv_sharing_target(current_layer_name, target_layer_name,
+                               static_forward_context):
+    error_msg = (f"Specified KV sharing target layer for {current_layer_name} "
+                 f"is not valid: target layer {target_layer_name} ")
+
+    if current_layer_name == target_layer_name:
+        raise ValueError(error_msg +
+                         "cannot be the same as the current layer.")
+
+    if target_layer_name not in static_forward_context:
+        from vllm.model_executor.models.utils import extract_layer_index
+
+        # If target layer name is not in the static fwd context, it means either
+        # a) the target layer does not come BEFORE the current layer, or
+        # b) the target layer is not an Attention layer that exists in the model
+        current_layer_idx = extract_layer_index(current_layer_name)
+        target_layer_idx = extract_layer_index(target_layer_name)
+        if current_layer_idx <= target_layer_idx:
+            raise ValueError(error_msg + "must come before the current layer.")
+        else:
+            raise ValueError(error_msg +
+                             "is not a valid Attention layer in the model.")
+
+    # Currently KV sharing is only supported between layers of the same type
+    target_layer_attn_type = static_forward_context[
+        target_layer_name].attn_type
+    expected = static_forward_context[current_layer_name].attn_type
+    if target_layer_attn_type != expected:
+        raise ValueError(
+            error_msg +
+            f"must be the same type as the current layer ({expected}).")

vllm/v1/worker/gpu_model_runner.py

@@ -59,8 +59,8 @@ from vllm.v1.worker.block_table import BlockTable
 from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
 from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
 
-from .utils import (gather_mm_placeholders, sanity_check_mm_encoder_outputs,
-                    scatter_mm_placeholders)
+from .utils import (gather_mm_placeholders, initialize_kv_cache_for_kv_sharing,
+                    sanity_check_mm_encoder_outputs, scatter_mm_placeholders)
 
 if TYPE_CHECKING:
     import xgrammar as xgr
@@ -276,6 +276,12 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                                         pin_memory=self.pin_memory)
         self.seq_lens_np = self.seq_lens_cpu.numpy()
 
+        # Layer pairings for cross-layer KV sharing.
+        # If an Attention layer `layer_name` is in the keys of this dict, it
+        # means this layer will perform attention using the keys and values
+        # from the KV cache of `shared_kv_cache_layers[layer_name]`.
+        self.shared_kv_cache_layers: dict[str, str] = {}
+
     def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> bool:
         """
         Update the order of requests in the batch based on the attention
@@ -2097,6 +2103,15 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                     # KV cache specs.
                     raise ValueError("Unknown KV cache spec type.")
 
+        # Setup `kv_cache_config` and `kv_caches` for models
+        # with cross-layer KV sharing
+        if self.shared_kv_cache_layers:
+            initialize_kv_cache_for_kv_sharing(
+                self.shared_kv_cache_layers,
+                kv_cache_config.kv_cache_groups,
+                kv_caches,
+            )
+
         if self.speculative_config and self.speculative_config.use_eagle():
             assert isinstance(self.drafter, EagleProposer)
             # validate all draft model layers belong to the same kv cache
@@ -2125,6 +2140,18 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         use_mla = self.vllm_config.model_config.use_mla
         kv_cache_spec: dict[str, KVCacheSpec] = {}
         for layer_name, attn_module in layers.items():
+            if (kv_tgt_layer :=
+                    attn_module.kv_sharing_target_layer_name) is not None:
+                # The layer doesn't need its own KV cache and will use that of
+                # the target layer. We skip creating a KVCacheSpec for it, so
+                # that KV cache management logic will act as this layer does
+                # not exist, and doesn't allocate KV cache for the layer. This
+                # enables the memory saving of cross-layer kv sharing, allowing
+                # a given amount of memory to accommodate longer context lengths
+                # or enable more requests to be processed simultaneously.
+                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
+                continue
+
             # TODO: Support other attention modules, e.g., cross-attention
             if attn_module.attn_type == AttentionType.DECODER:
                 if attn_module.sliding_window is not None:

vllm/v1/worker/tpu_model_runner.py

@@ -44,7 +44,8 @@ from vllm.v1.utils import bind_kv_cache
 from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
 from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
 
-from .utils import sanity_check_mm_encoder_outputs
+from .utils import (initialize_kv_cache_for_kv_sharing,
+                    sanity_check_mm_encoder_outputs)
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
@@ -238,6 +239,12 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         self.num_reqs_paddings = _get_req_paddings(
             min_req_size=MIN_NUM_SEQS, max_req_size=self.max_num_reqs)
 
+        # Layer pairings for cross-layer KV sharing.
+        # If an Attention layer `layer_name` is in the keys of this dict, it
+        # means this layer will perform attention using the keys and values
+        # from the KV cache of `shared_kv_cache_layers[layer_name]`.
+        self.shared_kv_cache_layers: dict[str, str] = {}
+
         # tensors for structured decoding
         self.grammar_bitmask_cpu = torch.zeros(
             (self.max_num_reqs, cdiv(self.vocab_size, 32)),
@@ -455,6 +462,18 @@ class TPUModelRunner(LoRAModelRunnerMixin):
         block_size = self.vllm_config.cache_config.block_size
         kv_cache_spec: dict[str, KVCacheSpec] = {}
         for layer_name, attn_module in layers.items():
+            if (kv_tgt_layer :=
+                    attn_module.kv_sharing_target_layer_name) is not None:
+                # The layer doesn't need its own KV cache and will use that of
+                # the target layer. We skip creating a KVCacheSpec for it, so
+                # that KV cache management logic will act as this layer does
+                # not exist, and doesn't allocate KV cache for the layer. This
+                # enables the memory saving of cross-layer kv sharing, allowing
+                # a given amount of memory to accommodate longer context lengths
+                # or enable more requests to be processed simultaneously.
+                self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
+                continue
+
             if attn_module.attn_type == AttentionType.DECODER:
                 if attn_module.sliding_window is not None:
                     kv_cache_spec[layer_name] = SlidingWindowSpec(
@@ -1376,6 +1395,15 @@ class TPUModelRunner(LoRAModelRunnerMixin):
                 else:
                     raise NotImplementedError
 
+        # Setup `kv_cache_config` and `kv_caches` for models
+        # with cross-layer KV sharing
+        if self.shared_kv_cache_layers:
+            initialize_kv_cache_for_kv_sharing(
+                self.shared_kv_cache_layers,
+                kv_cache_config.kv_cache_groups,
+                kv_caches,
+            )
+
         bind_kv_cache(
             kv_caches,
             self.vllm_config.compilation_config.static_forward_context,

vllm/v1/worker/utils.py

@@ -4,6 +4,8 @@ from typing import Optional
 
 import torch
 
+from vllm.v1.kv_cache_interface import KVCacheGroupSpec
+
 
 def sanity_check_mm_encoder_outputs(
     mm_embeddings: object,
@@ -73,3 +75,37 @@ def gather_mm_placeholders(
         return placeholders
 
     return placeholders[is_embed]
+
+
+def initialize_kv_cache_for_kv_sharing(
+    shared_kv_cache_layers: dict[str, str],
+    kv_cache_groups: list[KVCacheGroupSpec],
+    kv_caches: dict[str, torch.Tensor],
+) -> None:
+    """
+    Sets up KV cache sharing by reusing the allocated KV caches in `kv_caches`
+    for layers that do not allocate its own KV cache, based on the mapping in
+    `shared_kv_cache_layers`. Adds these layers to the corresponding KV cache
+    group, which is needed to ensure that attention metadata is assigned later.
+
+    Args:
+        shared_kv_cache_layers: Layer pairings for cross-layer KV sharing.
+            If an Attention layer `layer_name` is in the keys of this dict, it
+            means this layer will perform attention using the keys and values
+            from the KV cache of `shared_kv_cache_layers[layer_name]`.
+        kv_cache_groups: The KV cache groups of the model.
+        kv_caches: The allocated kv_caches with layer names as keys.
+            Note that layers in shared_kv_cache_layers.keys() are not
+            originally included as it only contains layers which have its own
+            KV cache allocation.
+    """
+    # Record index of KV cache group for each layer that allocates a KV cache.
+    layer_to_kv_cache_group_idx: dict[str, int] = {}
+    for i, kv_cache_group in enumerate(kv_cache_groups):
+        for layer_name in kv_cache_group.layer_names:
+            layer_to_kv_cache_group_idx[layer_name] = i
+
+    for layer_name, target_layer_name in shared_kv_cache_layers.items():
+        kv_caches[layer_name] = kv_caches[target_layer_name]
+        group_idx = layer_to_kv_cache_group_idx[target_layer_name]
+        kv_cache_groups[group_idx].layer_names.append(layer_name)