Add FlashInfer support (#2354)

This change adds support for FlashInfer. FlashInfer can be enabled using
`FLASH_INFER=1` and is currently only implemented in `FlashCausalLM`.
Since this functionality is currently only for testing, FlashInfer is
not installed anywhere yet.

The FlashInfer API is quite different from FlashAttention/vLLM in that
it requires more global bookkeeping:

* A wrapper class needs to be contstructed (which we just call *state*).
  Since this is fairly expensive (due to pinned host memory allocation),
  we only do this once in a FlashCausalLM instance or for each CUDA
  Graph size.
* Each model forward call needs to be wrapped in `begin_forward` and
  `end_forward`. This sets up data structures that can be reused for all
  calls to attention for that forward call.

When calling attention, we need access to the state object. To avoid
passing an argument down the call chain (which would require changes to
all models), we use a context variable.

Each model forward call is wrapped using a context manager that does all
the bookkeeping for such a call:

* Set the context variable to the forward call's state.
* Call `begin_forward` on the state.
* Yield.
* Call `end_forward` on the state.
* Reset the context variable.

We cannot use a single shared global variable for this, since e.g. CUDA
Graphs of different sizes each have their own state.

server/text_generation_server/layers/attention/common.py

@@ -1,10 +1,10 @@
 from dataclasses import dataclass
-from text_generation_server.models.globals import FLASH_DECODING
+from text_generation_server.models.globals import FLASH_DECODING, FLASH_INFER
 import torch
 from typing import Optional
 
 
-if FLASH_DECODING:
+if FLASH_DECODING or FLASH_INFER:
 
     @dataclass
     class Seqlen:

server/text_generation_server/layers/attention/cuda.py

@@ -1,6 +1,10 @@
 import torch
 from text_generation_server.utils.import_utils import SYSTEM
-from text_generation_server.models.globals import FLASH_DECODING, BLOCK_SIZE
+from text_generation_server.models.globals import (
+    FLASH_DECODING,
+    BLOCK_SIZE,
+    FLASH_INFER,
+)
 from text_generation_server.layers.attention import Seqlen
 from typing import Optional
 
@@ -23,7 +27,7 @@ def reshape_and_cache(
     value_cache: torch.Tensor,
     slots: torch.Tensor,
 ):
-    if FLASH_DECODING:
+    if FLASH_DECODING or FLASH_INFER:
         shape = key_cache.shape
         key_cache.view(-1, shape[-2], shape[-1])[slots] = key
         value_cache.view(-1, shape[-2], shape[-1])[slots] = value
@@ -72,7 +76,16 @@ def paged_attention(
     # V1 to avoid the overhead of reduction. Also, if the number of
     # sequences or heads is large, we use V1 since there is enough work
     # to parallelize.
-    if FLASH_DECODING:
+    if FLASH_INFER:
+        from text_generation_server.layers.attention.flash_infer import decode_state
+
+        return decode_state.get().forward(
+            query.contiguous(),
+            paged_kv_cache=(key_cache, value_cache),
+            logits_soft_cap=softcap,
+            sm_scale=softmax_scale,
+        )
+    elif FLASH_DECODING:
         max_q = 1
         max_k = max_s
         import flash_attn_2_cuda
@@ -206,7 +219,32 @@ except ImportError:
 
 SUPPORTS_WINDOWING = V2
 
-if V2:
+if FLASH_INFER:
+
+    def attention(
+        q,
+        k,
+        v,
+        cu_seqlens,
+        max_s,
+        softmax_scale,
+        window_size_left=-1,
+        causal=True,
+        softcap=0.0,
+    ):
+        from text_generation_server.layers.attention.flash_infer import prefill_state
+
+        return prefill_state.get().forward(
+            q,
+            k,
+            v,
+            causal=causal,
+            window_left=window_size_left,
+            logits_soft_cap=softcap,
+            sm_scale=softmax_scale,
+        )
+
+elif V2:
 
     def attention(
         q,

server/text_generation_server/layers/attention/flash_infer.py

@@ -0,0 +1,164 @@
+from typing import Optional
+from contextvars import ContextVar
+from contextlib import contextmanager
+
+import flashinfer
+import torch
+
+prefill_state: ContextVar[flashinfer.BatchPrefillWithRaggedKVCacheWrapper] = ContextVar(
+    "prefill_state"
+)
+
+decode_state: ContextVar[flashinfer.BatchDecodeWithPagedKVCacheWrapper] = ContextVar(
+    "decode_state"
+)
+
+workspace: Optional[torch.Tensor] = None
+
+
+def get_workspace(device):
+    """Get shared flashinfer workspace."""
+    global workspace
+    if workspace is None:
+        workspace = torch.empty(128 * 1024 * 1024, dtype=torch.uint8, device=device)
+    return workspace
+
+
+def create_prefill_state(
+    *,
+    device: torch.device,
+):
+    """Create a prefill state."""
+    workspace_buffer = get_workspace(device)
+    return flashinfer.BatchPrefillWithRaggedKVCacheWrapper(
+        workspace_buffer, kv_layout="NHD", use_cuda_graph=False
+    )
+
+
+@contextmanager
+def use_prefill_state(
+    *,
+    state: flashinfer.BatchPrefillWithRaggedKVCacheWrapper,
+    cu_seqlens: torch.Tensor,
+    num_heads: int,
+    num_kv_heads: int,
+    head_size: int,
+    query_dtype: str = "float16",
+):
+    """
+    Context manager to set the active flashinfer prefill state to the given
+    `state` and parameters. This state will be used by all calls to the
+    `attention` function while the context manager is active.
+    """
+
+    token = prefill_state.set(state)
+    try:
+        state.begin_forward(
+            qo_indptr=cu_seqlens,
+            kv_indptr=cu_seqlens,
+            num_qo_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=head_size,
+            q_data_type=query_dtype,
+        )
+        yield
+    finally:
+        state.end_forward()
+        if token is not None:
+            prefill_state.reset(token)
+
+
+def create_decode_state(
+    *,
+    device: torch.device,
+    num_heads: int,
+    num_kv_heads: int,
+):
+    """Create a decode state."""
+    workspace_buffer = get_workspace(device)
+    return flashinfer.BatchDecodeWithPagedKVCacheWrapper(
+        workspace_buffer,
+        kv_layout="NHD",
+        use_cuda_graph=False,
+        use_tensor_cores=num_heads // num_kv_heads > 4,
+    )
+
+
+def create_decode_state_cuda_graphs(
+    *,
+    device: torch.device,
+    block_tables: torch.Tensor,
+    block_tables_ptr: torch.Tensor,
+    last_page_len: torch.Tensor,
+    num_heads: int,
+    num_kv_heads: int,
+):
+    """
+    Create a decode state for use with CUDA Graphs. `block_tables`,
+    `block_tables_ptr`, and `last_page_len` are used in CUDA Graphs and are
+    therefore stored as part of the state.
+    """
+    workspace_buffer = get_workspace(device)
+    return flashinfer.BatchDecodeWithPagedKVCacheWrapper(
+        workspace_buffer,
+        kv_layout="NHD",
+        use_cuda_graph=True,
+        paged_kv_indices_buffer=block_tables,
+        paged_kv_indptr_buffer=block_tables_ptr,
+        paged_kv_last_page_len_buffer=last_page_len,
+        use_tensor_cores=num_heads // num_kv_heads > 4,
+    )
+
+
+@contextmanager
+def use_decode_state(
+    *,
+    state: flashinfer.BatchDecodeWithPagedKVCacheWrapper,
+    input_lengths: torch.Tensor,
+    block_tables: torch.Tensor,
+    num_heads: int,
+    num_kv_heads: int,
+    head_size: int,
+    page_size: int,
+    query_dtype: str = "float16",
+):
+    """
+    Context manager to set the active flashinfer decoding state to the given
+    `state` and parameters. This state will be used by all calls to the
+    `paged_attention` function while the context manager is active.
+    """
+    indptr = torch.zeros(
+        input_lengths.shape[0] + 1, device=input_lengths.device, dtype=torch.int32
+    )
+    # Round up to page size and then calculate the cumulative sum to get
+    # the indices into the block table.
+    torch.add(input_lengths, page_size - 1, out=indptr[1:])
+    indptr[1:].div_(page_size, rounding_mode="floor")
+    indptr[1:].cumsum_(-1)
+
+    # Get the lengths of the last page in a block.
+    last_page_len = torch.empty(
+        input_lengths.shape[0], dtype=torch.int32, device=input_lengths.device
+    )
+    torch.sub(input_lengths, 1, out=last_page_len)
+    last_page_len.remainder_(page_size)
+    last_page_len += 1
+
+    token = decode_state.set(state)
+
+    try:
+        state.begin_forward(
+            indptr=indptr,
+            indices=block_tables,
+            last_page_len=last_page_len,
+            num_qo_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=head_size,
+            page_size=page_size,
+            q_data_type=query_dtype,
+        )
+        yield
+    finally:
+        state.end_forward()
+        if token is not None:
+            decode_state.reset(token)

server/text_generation_server/models/flash_causal_lm.py

@@ -1,3 +1,4 @@
+from contextlib import nullcontext
 import math
 import os
 import time
@@ -15,7 +16,7 @@ from transformers import (
     AutoTokenizer,
     GenerationConfig,
 )
-from typing import Iterable, Optional, Tuple, List, Type, Dict
+from typing import Any, ContextManager, Iterable, Optional, Tuple, List, Type, Dict
 
 from text_generation_server.adapters import AdapterBatchData, AdapterBatchMetadata
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
@@ -40,6 +41,7 @@ from text_generation_server.pb import generate_pb2
 from text_generation_server.models.globals import (
     MEM_POOL,
     FLASH_DECODING,
+    FLASH_INFER,
     BLOCK_SIZE,
     CUDA_GRAPHS,
     get_adapter_to_index,
@@ -907,6 +909,7 @@ class FlashCausalLM(Model):
             config.sliding_window = None
 
         self.num_layers = config.num_hidden_layers
+        self.num_heads = config.num_attention_heads
         # Validation is done in the model itself
         if num_kv_heads is None:
             num_kv_heads = getattr(config, "num_key_value_heads", None)
@@ -935,6 +938,21 @@ class FlashCausalLM(Model):
         self.cuda_graphs = {}
         self.kv_cache = []
 
+        if FLASH_INFER:
+            from text_generation_server.layers.attention.flash_infer import (
+                create_prefill_state,
+                create_decode_state,
+            )
+
+            self.prefill_state = create_prefill_state(device=device)
+
+            if not CUDA_GRAPHS:
+                self.decode_state = create_decode_state(
+                    device=device,
+                    num_heads=self.num_heads,
+                    num_kv_heads=self.num_kv_heads,
+                )
+
         super().__init__(
             model_id=model_id,
             model=model,
@@ -972,7 +990,7 @@ class FlashCausalLM(Model):
         else:
             x = BLOCK_SIZE // element_size
 
-        if FLASH_DECODING:
+        if FLASH_DECODING or FLASH_INFER:
             self.kv_cache = [
                 (
                     torch.empty(
@@ -1044,38 +1062,66 @@ class FlashCausalLM(Model):
         graph = torch.cuda.CUDAGraph()
         self.cuda_graphs[bs]["graph"] = graph
 
+        if FLASH_INFER:
+            from text_generation_server.layers.attention.flash_infer import (
+                create_decode_state_cuda_graphs,
+            )
+
+            block_tables_ptr = torch.zeros(
+                bs + 1, dtype=torch.int32, device=self.device
+            )
+            last_page_len = torch.ones(bs, dtype=torch.int32, device=self.device)
+            state = create_decode_state_cuda_graphs(
+                device=input_ids.device,
+                block_tables=block_tables.view(-1),
+                block_tables_ptr=block_tables_ptr,
+                last_page_len=last_page_len,
+                num_heads=self.num_heads,
+                num_kv_heads=self.num_kv_heads,
+            )
+            self.cuda_graphs[bs]["state"] = state
+        else:
+            state = None
+
         torch.cuda.synchronize()
         # Run once outside to warmup
-        self.model.forward(
+        with self._forward_context(
-            input_ids=input_ids,
-            position_ids=position_ids,
-            cu_seqlen_prefill=None,
-            kv_cache=self.kv_cache,
             block_tables=block_tables,
-            slots=slots,
+            cu_seqlen_prefill=None,
-            input_lengths=input_lengths_,
+            input_lengths=input_lengths,
-            max_s=max_s,
+            state=state,
-            prefill_cache_indices=None,
+        ):
-            lm_head_indices=None,
+            self.model.forward(
-        )
-        torch.cuda.synchronize()
-
-        with torch.cuda.graph(graph, pool=MEM_POOL):
-            input_lengths = Seqlen(input_lengths=input_lengths)
-            logits, speculative_logits = self.model.forward(
                 input_ids=input_ids,
                 position_ids=position_ids,
                 cu_seqlen_prefill=None,
                 kv_cache=self.kv_cache,
                 block_tables=block_tables,
                 slots=slots,
-                input_lengths=input_lengths,
+                input_lengths=input_lengths_,
                 max_s=max_s,
                 prefill_cache_indices=None,
                 lm_head_indices=None,
             )
-            self.cuda_graphs[bs]["logits"] = logits
+
-            self.cuda_graphs[bs]["speculative_logits"] = speculative_logits
+            torch.cuda.synchronize()
+
+            with torch.cuda.graph(graph, pool=MEM_POOL):
+                input_lengths = Seqlen(input_lengths=input_lengths)
+                logits, speculative_logits = self.model.forward(
+                    input_ids=input_ids,
+                    position_ids=position_ids,
+                    cu_seqlen_prefill=None,
+                    kv_cache=self.kv_cache,
+                    block_tables=block_tables,
+                    slots=slots,
+                    input_lengths=input_lengths,
+                    max_s=max_s,
+                    prefill_cache_indices=None,
+                    lm_head_indices=None,
+                )
+                self.cuda_graphs[bs]["logits"] = logits
+                self.cuda_graphs[bs]["speculative_logits"] = speculative_logits
         torch.cuda.synchronize()
 
     def warmup(self, batch: FlashCausalLMBatch):
@@ -1295,23 +1341,28 @@ class FlashCausalLM(Model):
             cuda_graph = None
 
         if cu_seqlen_prefill is not None or cuda_graph is None:
-            input_lengths = Seqlen(input_lengths=input_lengths)
+            with self._forward_context(
-            logits, speculative_logits = self.model.forward(
-                input_ids=input_ids,
-                position_ids=position_ids,
-                cu_seqlen_prefill=cu_seqlen_prefill,
-                kv_cache=kv_cache,
                 block_tables=block_tables,
-                slots=slots,
+                cu_seqlen_prefill=cu_seqlen_prefill,
                 input_lengths=input_lengths,
-                max_s=max_s,
+            ):
-                prefill_cache_indices=batch.prefill_cache_indices,
+                input_lengths = Seqlen(input_lengths=input_lengths)
-                lm_head_indices=lm_head_indices,
+                logits, speculative_logits = self.model.forward(
-                adapter_data=adapter_data,
+                    input_ids=input_ids,
-            )
+                    position_ids=position_ids,
-            if batch.prefill_cache_indices is not None:
+                    cu_seqlen_prefill=cu_seqlen_prefill,
-                batch.prefill_cache_indices = None
+                    kv_cache=kv_cache,
-            return logits, speculative_logits
+                    block_tables=block_tables,
+                    slots=slots,
+                    input_lengths=input_lengths,
+                    max_s=max_s,
+                    prefill_cache_indices=batch.prefill_cache_indices,
+                    lm_head_indices=lm_head_indices,
+                    adapter_data=adapter_data,
+                )
+                if batch.prefill_cache_indices is not None:
+                    batch.prefill_cache_indices = None
+                return logits, speculative_logits
 
         # Copy inputs to the static inputs of the cuda graph
         # Static inputs are potentially padded
@@ -1325,8 +1376,16 @@ class FlashCausalLM(Model):
         cuda_graph["input_lengths"].zero_()
         cuda_graph["input_lengths"][: input_lengths.shape[0]] = input_lengths
 
-        # Replay the graph
+        state = cuda_graph.get("state")
-        cuda_graph["graph"].replay()
+        with self._forward_context(
+            block_tables=block_tables,
+            cu_seqlen_prefill=None,
+            input_lengths=input_lengths,
+            state=state,
+        ):
+            # Replay the graph
+            cuda_graph["graph"].replay()
+
         # Slice output to the correct shape
         speculative_logits = (
             cuda_graph["speculative_logits"][:bs]
@@ -1698,3 +1757,39 @@ class FlashCausalLM(Model):
         forward_ns = start_decode - start
         decode_ns = time.time_ns() - start_decode
         return generations, batch, (forward_ns, decode_ns)
+
+    def _forward_context(
+        self,
+        *,
+        block_tables: torch.Tensor,
+        cu_seqlen_prefill: Optional[torch.Tensor],
+        input_lengths: torch.Tensor,
+        state: Optional[Any] = None,
+    ) -> ContextManager:
+        if not FLASH_INFER:
+            return nullcontext()
+
+        from text_generation_server.layers.attention.flash_infer import (
+            use_decode_state,
+            use_prefill_state,
+        )
+
+        if cu_seqlen_prefill is not None:
+            return use_prefill_state(
+                state=state if state is not None else self.prefill_state,
+                cu_seqlens=cu_seqlen_prefill,
+                num_heads=self.num_heads,
+                num_kv_heads=self.num_kv_heads,
+                head_size=self.head_size,
+            )
+        else:
+            assert input_lengths is not None
+            return use_decode_state(
+                state=state if state is not None else self.decode_state,
+                input_lengths=input_lengths,
+                block_tables=block_tables.view(-1),
+                num_heads=self.num_heads,
+                num_kv_heads=self.num_kv_heads,
+                head_size=self.head_size,
+                page_size=BLOCK_SIZE,
+            )

server/text_generation_server/models/globals.py

@@ -5,6 +5,10 @@ from typing import Dict, Optional
 
 from text_generation_server.utils.log import log_master
 
+FLASH_INFER = os.getenv("FLASH_INFER") in {"1", "true", "True"}
+if FLASH_INFER:
+    log_master(logger.info, "Using FLASH_INFER")
+
 MEM_POOL = torch.cuda.graph_pool_handle() if torch.cuda.is_available() else None
 # This is overridden by the cli
 FLASH_DECODING = os.getenv("FLASH_DECODING") in {"1", "true", "True"}
@@ -12,6 +16,7 @@ BLOCK_SIZE: int = 256 if FLASH_DECODING else 16
 if FLASH_DECODING:
     log_master(logger.info, "Using FLASH_DECODING")
 
+
 cuda_graphs = os.getenv("CUDA_GRAPHS")
 if cuda_graphs is not None:
     try: