refactor EAGLE 2 (#3269)

Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: merrymercy <lianminzheng@gmail.com>
Co-authored-by: Ying1123 <sqy1415@gmail.com>

examples/runtime/engine/EAGLE_offline_batch_inference.py

@@ -21,6 +21,7 @@ def main():
         speculative_num_steps=3,
         speculative_eagle_topk=4,
         speculative_num_draft_tokens=16,
+        cuda_graph_max_bs=8,
     )
 
     outputs = llm.generate(prompts, sampling_params)

python/sglang/srt/layers/attention/flashinfer_backend.py

@@ -10,6 +10,7 @@ Each backend supports two operators: extend (i.e. prefill with cached prefix) an
 import os
 from dataclasses import dataclass
 from enum import Enum, auto
+from functools import partial
 from typing import TYPE_CHECKING, List, Optional, Union
 
 import torch
@@ -34,6 +35,7 @@ if is_flashinfer_available():
         BatchPrefillWithRaggedKVCacheWrapper,
     )
     from flashinfer.cascade import merge_state
+    from flashinfer.decode import PosEncodingMode
 
 
 class WrapperDispatch(Enum):
@@ -53,10 +55,19 @@ class PrefillMetadata:
     extend_no_prefix: bool
 
 
+# Reuse this workspace buffer across all flashinfer wrappers
+global_workspace_buffer = None
+
+
 class FlashInferAttnBackend(AttentionBackend):
     """Flashinfer attention kernels."""
 
-    def __init__(self, model_runner: ModelRunner):
+    def __init__(
+        self,
+        model_runner: ModelRunner,
+        skip_prefill: bool = False,
+        kv_indptr_buf: Optional[torch.Tensor] = None,
+    ):
         super().__init__()
 
         # Parse constants
@@ -69,6 +80,7 @@ class FlashInferAttnBackend(AttentionBackend):
             ),
         )
         self.max_context_len = model_runner.model_config.context_len
+        self.skip_prefill = skip_prefill
 
         assert not (
             model_runner.sliding_window_size is not None
@@ -90,16 +102,26 @@ class FlashInferAttnBackend(AttentionBackend):
             global_config.flashinfer_workspace_size = 512 * 1024 * 1024
 
         # Allocate buffers
-        self.workspace_buffer = torch.empty(
+        global global_workspace_buffer
+        if global_workspace_buffer is None:
+            global_workspace_buffer = torch.empty(
                 global_config.flashinfer_workspace_size,
                 dtype=torch.uint8,
                 device=model_runner.device,
             )
+        self.workspace_buffer = global_workspace_buffer
         max_bs = model_runner.req_to_token_pool.size
+        if kv_indptr_buf is None:
             self.kv_indptr = [
-            torch.zeros((max_bs + 1,), dtype=torch.int32, device=model_runner.device)
+                torch.zeros(
+                    (max_bs + 1,), dtype=torch.int32, device=model_runner.device
+                )
                 for _ in range(self.num_wrappers)
             ]
+        else:
+            assert self.num_wrappers == 1
+            self.kv_indptr = [kv_indptr_buf]
+
         self.kv_last_page_len = torch.ones(
             (max_bs,), dtype=torch.int32, device=model_runner.device
         )
@@ -122,8 +144,12 @@ class FlashInferAttnBackend(AttentionBackend):
         self.prefill_wrappers_verify = []
         self.decode_wrappers = []
         for _ in range(self.num_wrappers):
+            if not skip_prefill:
                 self.prefill_wrappers_paged.append(
-                BatchPrefillWithPagedKVCacheWrapper(self.workspace_buffer, "NHD")
+                    BatchPrefillWithPagedKVCacheWrapper(
+                        self.workspace_buffer,
+                        "NHD",
+                    )
                 )
                 self.prefill_wrappers_verify.append(
                     BatchPrefillWithPagedKVCacheWrapper(self.workspace_buffer, "NHD")
@@ -137,10 +163,11 @@ class FlashInferAttnBackend(AttentionBackend):
             )
 
         # Create indices updater
-        self.indices_updater_decode = FlashInferIndicesUpdaterDecode(model_runner, self)
+        if not skip_prefill:
             self.indices_updater_prefill = FlashInferIndicesUpdaterPrefill(
                 model_runner, self
             )
+        self.indices_updater_decode = FlashInferIndicesUpdaterDecode(model_runner, self)
 
         # Other metadata
         self.forward_metadata: Union[PrefillMetadata, DecodeMetadata] = None
@@ -211,16 +238,23 @@ class FlashInferAttnBackend(AttentionBackend):
                 self.prefill_wrappers_paged, use_ragged, extend_no_prefix
             )
 
-    def init_cuda_graph_state(self, max_bs: int):
+    def init_cuda_graph_state(
+        self, max_bs: int, kv_indices_buf: Optional[torch.Tensor] = None
+    ):
+        if kv_indices_buf is None:
             cuda_graph_kv_indices = torch.zeros(
                 (max_bs * self.max_context_len,),
                 dtype=torch.int32,
                 device="cuda",
             )
+        else:
+            cuda_graph_kv_indices = kv_indices_buf
+
         self.cuda_graph_kv_indices = [cuda_graph_kv_indices] + [
             cuda_graph_kv_indices.clone() for _ in range(self.num_wrappers - 1)
         ]
 
+        if not self.skip_prefill:
             self.cuda_graph_custom_mask = torch.zeros(
                 (max_bs * self.max_context_len),
                 dtype=torch.uint8,
@@ -602,11 +636,8 @@ class FlashInferIndicesUpdaterDecode:
                 self.req_to_token.shape[1],
             )
         else:
-            bs, kv_indices, kv_indptr = spec_info.generate_attn_arg_decode(
-                req_pool_indices,
-                paged_kernel_lens,
-                self.req_to_token,
-            )
+            kv_indptr, kv_indices = spec_info.kv_indptr, spec_info.kv_indices
+            bs = kv_indptr.shape[0] - 1
 
         wrapper.end_forward()
         wrapper.begin_forward(
@@ -854,6 +885,132 @@ class FlashInferIndicesUpdaterPrefill:
         )
 
 
+class FlashInferMultiStepDraftBackend:
+    """
+    Wrap multiple flashinfer attention backends as one for multiple consecutive
+    draft decoding steps.
+    """
+
+    def __init__(
+        self,
+        model_runner: ModelRunner,
+        topk: int,
+        speculative_num_steps: int,
+    ):
+        from sglang.srt.speculative.eagle_utils import generate_draft_decode_kv_indices
+
+        self.topk = topk
+        self.speculative_num_steps = speculative_num_steps
+        self.generate_draft_decode_kv_indices = generate_draft_decode_kv_indices
+        max_bs = model_runner.req_to_token_pool.size
+        self.kv_indptr = torch.zeros(
+            (
+                self.speculative_num_steps,
+                max_bs + 1,
+            ),
+            dtype=torch.int32,
+            device=model_runner.device,
+        )
+        self.attn_backends = []
+        for i in range(self.speculative_num_steps):
+            self.attn_backends.append(
+                FlashInferAttnBackend(
+                    model_runner,
+                    skip_prefill=True,
+                    kv_indptr_buf=self.kv_indptr[i],
+                )
+            )
+        self.max_context_len = self.attn_backends[0].max_context_len
+        # Cached variables for generate_draft_decode_kv_indices
+        self.pool_len = model_runner.req_to_token_pool.req_to_token.shape[1]
+        self.kv_indptr_stride = self.kv_indptr.shape[1]
+
+    def common_template(self, forward_batch: ForwardBatch, call_fn: int):
+        num_seqs = forward_batch.batch_size
+        bs = self.topk * num_seqs
+        seq_lens_sum = forward_batch.seq_lens_sum
+        self.generate_draft_decode_kv_indices[
+            (self.speculative_num_steps, num_seqs, self.topk)
+        ](
+            forward_batch.req_pool_indices,
+            forward_batch.req_to_token_pool.req_to_token,
+            forward_batch.seq_lens,
+            self.cuda_graph_kv_indices,
+            self.kv_indptr,
+            forward_batch.positions,
+            num_seqs,
+            self.topk,
+            self.pool_len,
+            self.kv_indptr_stride,
+            self.kv_indptr.shape[1],
+            triton.next_power_of_2(num_seqs),
+            triton.next_power_of_2(self.speculative_num_steps),
+            triton.next_power_of_2(bs),
+        )
+        for i in range(self.speculative_num_steps):
+            forward_batch.spec_info.kv_indptr = self.kv_indptr[i, : bs + 1]
+            forward_batch.spec_info.kv_indices = self.cuda_graph_kv_indices[i][
+                : seq_lens_sum * self.topk + bs * (i + 1)
+            ]
+            call_fn(i, forward_batch)
+
+    def init_forward_metadata(self, forward_batch: ForwardBatch):
+        def call_fn(i, forward_batch):
+            forward_batch.spec_info.kv_indptr = (
+                forward_batch.spec_info.kv_indptr.clone()
+            )
+            forward_batch.spec_info.kv_indices = (
+                forward_batch.spec_info.kv_indices.clone()
+            )
+            self.attn_backends[i].init_forward_metadata(forward_batch)
+
+        self.common_template(forward_batch, call_fn)
+
+    def init_cuda_graph_state(self, max_bs: int):
+        self.cuda_graph_kv_indices = torch.zeros(
+            (self.speculative_num_steps, max_bs * self.max_context_len),
+            dtype=torch.int32,
+            device="cuda",
+        )
+        self.kv_indptr_stride = self.cuda_graph_kv_indices.shape[1]
+        for i in range(self.speculative_num_steps):
+            self.attn_backends[i].init_cuda_graph_state(
+                max_bs, kv_indices_buf=self.cuda_graph_kv_indices[i]
+            )
+
+    def init_forward_metadata_capture_cuda_graph(self, forward_batch: ForwardBatch):
+        def call_fn(i, forward_batch):
+            self.attn_backends[i].init_forward_metadata_capture_cuda_graph(
+                forward_batch.batch_size,
+                forward_batch.batch_size * self.topk,
+                forward_batch.req_pool_indices,
+                forward_batch.seq_lens,
+                encoder_lens=None,
+                forward_mode=ForwardMode.DECODE,
+                spec_info=forward_batch.spec_info,
+            )
+            decode_wrapper = self.attn_backends[i].decode_cuda_graph_metadata[
+                forward_batch.batch_size
+            ][0]
+            decode_wrapper.begin_forward = partial(fast_decode_plan, decode_wrapper)
+
+        self.common_template(forward_batch, call_fn)
+
+    def init_forward_metadata_replay_cuda_graph(self, forward_batch):
+        def call_fn(i, forward_batch):
+            self.attn_backends[i].init_forward_metadata_replay_cuda_graph(
+                forward_batch.batch_size,
+                forward_batch.req_pool_indices,
+                forward_batch.seq_lens,
+                seq_lens_sum=-1,
+                encoder_lens=None,
+                forward_mode=ForwardMode.DECODE,
+                spec_info=forward_batch.spec_info,
+            )
+
+        self.common_template(forward_batch, call_fn)
+
+
 @triton.jit
 def create_flashinfer_kv_indices_triton(
     req_to_token_ptr,  # [max_batch, max_context_len]
@@ -937,3 +1094,105 @@ def should_use_tensor_core(
         return gqa_group_size > 4
     else:
         return False
+
+
+def fast_decode_plan(
+    self,
+    indptr: torch.Tensor,
+    indices: torch.Tensor,
+    last_page_len: torch.Tensor,
+    num_qo_heads: int,
+    num_kv_heads: int,
+    head_dim: int,
+    page_size: int,
+    pos_encoding_mode: str = "NONE",
+    window_left: int = -1,
+    logits_soft_cap: Optional[float] = None,
+    data_type: Union[str, torch.dtype] = "float16",
+    q_data_type: Optional[Union[str, torch.dtype]] = None,
+    sm_scale: Optional[float] = None,
+    rope_scale: Optional[float] = None,
+    rope_theta: Optional[float] = None,
+) -> None:
+    """A faster version of BatchDecodeWithPagedKVCacheWrapper::plan used for FlashInferMultiStepDraftBackend."""
+    batch_size = len(last_page_len)
+    if logits_soft_cap is None:
+        logits_soft_cap = 0.0
+    if self.is_cuda_graph_enabled:
+        if batch_size != self._fixed_batch_size:
+            raise ValueError(
+                "The batch size should be fixed in cudagraph mode, the runtime batch size {} "
+                " mismatches the batch size set during initialization {}".format(
+                    batch_size, self._fixed_batch_size
+                )
+            )
+        if len(indices) > len(self._paged_kv_indices_buf):
+            raise ValueError(
+                "The size of indices should be less than or equal to the allocated buffer"
+            )
+    else:
+        self._paged_kv_indptr_buf = indptr
+        self._paged_kv_indices_buf = indices
+        self._paged_kv_last_page_len_buf = last_page_len
+    # NOTE(Zihao): the following tensors acts as placeholder to pass dtype info
+    if not q_data_type:
+        q_data_type = data_type
+    if not hasattr(self, "empty_q_data"):
+        self.empty_q_data = torch.empty(
+            0,
+            dtype=(
+                getattr(torch, q_data_type)
+                if isinstance(q_data_type, str)
+                else q_data_type
+            ),
+        )
+        self.empty_kv_cache = torch.empty(
+            0,
+            dtype=(
+                getattr(torch, data_type) if isinstance(data_type, str) else data_type
+            ),
+        )
+        self.last_page_len = torch.ones(32768, dtype=torch.int32)
+    empty_q_data = self.empty_q_data
+    empty_kv_cache = self.empty_kv_cache
+    if self.use_tensor_cores:
+        if not self.is_cuda_graph_enabled:
+            # when not using cudagraph, we need to create the indptr buffer, otherwise
+            # the buffer is already created during initialization
+            self._qo_indptr_buf = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=indptr.device
+            )
+        self._wrapper.plan(
+            self._float_workspace_buffer,
+            self._int_workspace_buffer,
+            self._qo_indptr_buf,
+            indptr,
+            batch_size,
+            num_qo_heads,
+            num_kv_heads,
+            head_dim,
+            page_size,
+            empty_q_data,
+        )
+    else:
+        self._wrapper.plan(
+            self._float_workspace_buffer,
+            self._int_workspace_buffer,
+            indptr,
+            self.last_page_len,
+            batch_size,
+            num_qo_heads,
+            num_kv_heads,
+            head_dim,
+            page_size,
+            PosEncodingMode[pos_encoding_mode].value,
+            logits_soft_cap,
+            empty_q_data,
+            empty_kv_cache,
+        )
+    self._pos_encoding_mode = pos_encoding_mode
+    self._window_left = window_left
+    self._logits_soft_cap = logits_soft_cap
+    self._sm_scale = sm_scale
+    self._rope_scale = rope_scale
+    self._rope_theta = rope_theta

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -103,69 +103,75 @@ def set_torch_compile_config():
         torch._dynamo.config.cache_size_limit = 1024
 
 
-class CudaGraphRunner:
-    """A CudaGraphRunner runs the forward pass of a model with cuda graph and torch.compile."""
-
-    def __init__(self, model_runner: "ModelRunner"):
-        # Parse args
-        self.model_runner = model_runner
-        self.graphs = {}
-        self.input_buffers = {}
-        self.output_buffers = {}
-        self.flashinfer_handlers = {}
-        self.graph_memory_pool = None
-        self.use_torch_compile = model_runner.server_args.enable_torch_compile
-        self.disable_padding = model_runner.server_args.disable_cuda_graph_padding
-        self.is_encoder_decoder = self.model_runner.model_config.is_encoder_decoder
-        self.enable_dp_attention = self.model_runner.server_args.enable_dp_attention
-        self.tp_size = self.model_runner.tp_size
-        self.dp_size = self.model_runner.server_args.dp_size
-
-        # Batch sizes to capture
-        self.capture_bs = self.model_runner.server_args.cuda_graph_bs
-        if self.capture_bs is None:
-            if model_runner.server_args.disable_cuda_graph_padding:
-                self.capture_bs = list(range(1, 33)) + [64, 128]
+def get_batch_sizes_to_capture(model_runner: ModelRunner):
+    server_args = model_runner.server_args
+    capture_bs = server_args.cuda_graph_bs
+    if capture_bs is None:
+        if server_args.disable_cuda_graph_padding:
+            capture_bs = list(range(1, 33)) + [64, 128]
         else:
-                self.capture_bs = [1, 2, 4] + [i * 8 for i in range(1, 21)]
-
-        if max(self.capture_bs) > model_runner.req_to_token_pool.size:
+            capture_bs = [1, 2, 4] + [i * 8 for i in range(1, 21)]
+    if max(capture_bs) > model_runner.req_to_token_pool.size:
         # In some case (e.g., with a small GPU or --max-running-requests), the #max-running-requests
         # is very samll. We add more values here to make sure we capture the maximum bs.
-            self.capture_bs = list(
+        capture_bs = list(
             sorted(
                 set(
-                        self.capture_bs
+                    capture_bs
                     + [model_runner.req_to_token_pool.size - 1]
                     + [model_runner.req_to_token_pool.size]
                 )
             )
         )
-
-        self.capture_bs = [
+    capture_bs = [
         bs
-            for bs in self.capture_bs
+        for bs in capture_bs
         if bs <= model_runner.req_to_token_pool.size
-            and bs <= model_runner.server_args.cuda_graph_max_bs
-        ]
-
-        self.compile_bs = (
-            [
-                bs
-                for bs in self.capture_bs
-                if bs <= self.model_runner.server_args.torch_compile_max_bs
+        and bs <= server_args.cuda_graph_max_bs
     ]
-            if self.use_torch_compile
+    compile_bs = (
+        [bs for bs in capture_bs if bs <= server_args.torch_compile_max_bs]
+        if server_args.enable_torch_compile
         else []
     )
+    return capture_bs, compile_bs
+
+
+# Reuse this memory pool across all cuda graph runners.
+global_graph_memory_pool = None
+
+
+def get_global_graph_memory_pool():
+    return global_graph_memory_pool
+
 
+def set_global_graph_memory_pool(val):
+    global global_graph_memory_pool
+    global_graph_memory_pool = val
+
+
+class CudaGraphRunner:
+    """A CudaGraphRunner runs the forward pass of a model with cuda graph and torch.compile."""
+
+    def __init__(self, model_runner: ModelRunner):
+        # Parse args
+        self.model_runner = model_runner
+        self.graphs = {}
+        self.output_buffers = {}
+        self.enable_torch_compile = model_runner.server_args.enable_torch_compile
+        self.disable_padding = model_runner.server_args.disable_cuda_graph_padding
+        self.is_encoder_decoder = model_runner.model_config.is_encoder_decoder
+        self.enable_dp_attention = model_runner.server_args.enable_dp_attention
+        self.tp_size = model_runner.server_args.tp_size
+        self.dp_size = model_runner.server_args.dp_size
+
+        # Batch sizes to capture
+        self.capture_bs, self.compile_bs = get_batch_sizes_to_capture(model_runner)
         self.capture_forward_mode = ForwardMode.DECODE
         self.num_tokens_per_bs = 1
         if model_runner.spec_algorithm.is_eagle():
             if self.model_runner.is_draft_worker:
-                self.num_tokens_per_bs = (
-                    self.model_runner.server_args.speculative_eagle_topk
-                )
+                raise RuntimeError("This should not happen")
             else:
                 self.capture_forward_mode = ForwardMode.TARGET_VERIFY
                 self.num_tokens_per_bs = (
@@ -182,10 +188,10 @@ class CudaGraphRunner:
         # FIXME(lsyin): leave it here for now, I don't know whether it is necessary
         self.encoder_len_fill_value = 0
 
-        if self.use_torch_compile:
+        if self.enable_torch_compile:
             set_torch_compile_config()
 
-        # Common inputs
+        # Graph inputs
         with torch.device("cuda"):
             self.input_ids = torch.zeros((self.max_num_token,), dtype=torch.int64)
             self.req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int32)
@@ -301,7 +307,7 @@ class CudaGraphRunner:
         stream = self.stream
         num_tokens = bs * self.num_tokens_per_bs
 
-        # Common inputs
+        # Graph inputs
         input_ids = self.input_ids[:num_tokens]
         req_pool_indices = self.req_pool_indices[:bs]
         seq_lens = self.seq_lens[:bs]
@@ -320,7 +326,7 @@ class CudaGraphRunner:
             global_num_tokens = None
             gathered_buffer = None
 
-        spec_info = self.get_spec_info(num_tokens, positions)
+        spec_info = self.get_spec_info(num_tokens)
 
         forward_batch = ForwardBatch(
             forward_mode=self.capture_forward_mode,
@@ -335,7 +341,6 @@ class CudaGraphRunner:
             seq_lens_sum=seq_lens.sum(),
             encoder_lens=encoder_lens,
             return_logprob=False,
-            top_logprobs_nums=[0] * bs,
             positions=positions,
             global_num_tokens=global_num_tokens,
             gathered_buffer=gathered_buffer,
@@ -375,13 +380,14 @@ class CudaGraphRunner:
         torch.cuda.synchronize()
         self.model_runner.tp_group.barrier()
 
-        with torch.cuda.graph(graph, pool=self.graph_memory_pool, stream=stream):
+        global global_graph_memory_pool
+        with torch.cuda.graph(graph, pool=global_graph_memory_pool, stream=stream):
             out = run_once()
 
         torch.cuda.synchronize()
         self.model_runner.tp_group.barrier()
 
-        self.graph_memory_pool = graph.pool()
+        global_graph_memory_pool = graph.pool()
         return graph, out
 
     def replay(self, forward_batch: ForwardBatch):
@@ -439,35 +445,26 @@ class CudaGraphRunner:
         )
         return logits_output
 
-    def get_spec_info(self, num_tokens: int, positions: torch.Tensor):
+    def get_spec_info(self, num_tokens: int):
         spec_info = None
         if self.model_runner.spec_algorithm.is_eagle():
-            from sglang.srt.speculative.eagle_utils import (
-                EAGLEDraftInput,
-                EagleVerifyInput,
-            )
+            from sglang.srt.speculative.eagle_utils import EagleVerifyInput
 
             if self.model_runner.is_draft_worker:
-                spec_info = EAGLEDraftInput()
-                spec_info.load_server_args(self.model_runner.server_args)
-                spec_info.hidden_states = self.hidden_states[:num_tokens]
-                spec_info.positions = positions
-                spec_info.capture_hidden_mode = CaptureHiddenMode.FULL
+                raise RuntimeError("This should not happen.")
             else:
                 spec_info = EagleVerifyInput(
-                    None,
-                    None,
-                    None,
-                    None,
-                    None,
-                    None,
-                    self.model_runner.server_args.speculative_num_draft_tokens,
-                )
-                spec_info.custom_mask = torch.zeros(
+                    draft_token=None,
+                    custom_mask=torch.zeros(
                         (num_tokens * self.model_runner.model_config.context_len),
                         dtype=torch.bool,
                         device="cuda",
+                    ),
+                    positions=None,
+                    retrive_index=None,
+                    retrive_cum_len=None,
+                    draft_token_num=self.model_runner.server_args.speculative_num_draft_tokens,
+                    capture_hidden_mode=CaptureHiddenMode.FULL,
                 )
-                spec_info.capture_hidden_mode = CaptureHiddenMode.FULL
 
         return spec_info

python/sglang/srt/model_executor/forward_batch_info.py

@@ -197,64 +197,6 @@ class ForwardBatch:
     # For Qwen2-VL
     mrope_positions: torch.Tensor = None
 
-    def compute_mrope_positions(
-        self, model_runner: ModelRunner, batch: ModelWorkerBatch
-    ):
-        device = model_runner.device
-        hf_config = model_runner.model_config.hf_config
-        mrope_positions_list = [None] * self.seq_lens.shape[0]
-        if self.forward_mode.is_decode():
-            for i, _ in enumerate(mrope_positions_list):
-                mrope_position_delta = (
-                    0
-                    if batch.image_inputs[i] is None
-                    else batch.image_inputs[i].mrope_position_delta
-                )
-                mrope_positions_list[i] = MRotaryEmbedding.get_next_input_positions(
-                    mrope_position_delta,
-                    int(self.seq_lens[i]) - 1,
-                    int(self.seq_lens[i]),
-                )
-        elif self.forward_mode.is_extend():
-            extend_start_loc_cpu = self.extend_start_loc.cpu().numpy()
-            for i, image_inputs in enumerate(batch.image_inputs):
-                extend_start_loc, extend_seq_len, extend_prefix_len = (
-                    extend_start_loc_cpu[i],
-                    batch.extend_seq_lens[i],
-                    batch.extend_prefix_lens[i],
-                )
-                if image_inputs is None:
-                    # text only
-                    mrope_positions = [
-                        [
-                            pos
-                            for pos in range(
-                                extend_prefix_len, extend_prefix_len + extend_seq_len
-                            )
-                        ]
-                    ] * 3
-                else:
-                    # TODO: current qwen2-vl do not support radix cache since mrope position calculation
-                    mrope_positions, mrope_position_delta = (
-                        MRotaryEmbedding.get_input_positions(
-                            input_tokens=self.input_ids[
-                                extend_start_loc : extend_start_loc + extend_seq_len
-                            ],
-                            image_grid_thw=image_inputs.image_grid_thws,
-                            vision_start_token_id=hf_config.vision_start_token_id,
-                            spatial_merge_size=hf_config.vision_config.spatial_merge_size,
-                            context_len=0,
-                        )
-                    )
-                    batch.image_inputs[i].mrope_position_delta = mrope_position_delta
-                mrope_positions_list[i] = mrope_positions
-
-        self.mrope_positions = torch.concat(
-            [torch.tensor(pos, device=device) for pos in mrope_positions_list],
-            axis=1,
-        )
-        self.mrope_positions = self.mrope_positions.to(torch.int64)
-
     @classmethod
     def init_new(
         cls,
@@ -337,7 +279,7 @@ class ForwardBatch:
             ret.extend_logprob_start_lens_cpu = batch.extend_logprob_start_lens
 
         if model_runner.model_is_mrope:
-            ret.compute_mrope_positions(model_runner, batch)
+            ret._compute_mrope_positions(model_runner, batch)
 
         # Init lora information
         if model_runner.server_args.lora_paths is not None:
@@ -345,6 +287,63 @@ class ForwardBatch:
 
         return ret
 
+    def _compute_mrope_positions(
+        self, model_runner: ModelRunner, batch: ModelWorkerBatch
+    ):
+        device = model_runner.device
+        hf_config = model_runner.model_config.hf_config
+        mrope_positions_list = [None] * self.seq_lens.shape[0]
+        if self.forward_mode.is_decode():
+            for i, _ in enumerate(mrope_positions_list):
+                mrope_position_delta = (
+                    0
+                    if batch.image_inputs[i] is None
+                    else batch.image_inputs[i].mrope_position_delta
+                )
+                mrope_positions_list[i] = MRotaryEmbedding.get_next_input_positions(
+                    mrope_position_delta,
+                    int(self.seq_lens[i]) - 1,
+                    int(self.seq_lens[i]),
+                )
+        elif self.forward_mode.is_extend():
+            extend_start_loc_cpu = self.extend_start_loc.cpu().numpy()
+            for i, image_inputs in enumerate(batch.image_inputs):
+                extend_start_loc, extend_seq_len, extend_prefix_len = (
+                    extend_start_loc_cpu[i],
+                    batch.extend_seq_lens[i],
+                    batch.extend_prefix_lens[i],
+                )
+                if image_inputs is None:
+                    # text only
+                    mrope_positions = [
+                        [
+                            pos
+                            for pos in range(
+                                extend_prefix_len, extend_prefix_len + extend_seq_len
+                            )
+                        ]
+                    ] * 3
+                else:
+                    # TODO: current qwen2-vl do not support radix cache since mrope position calculation
+                    mrope_positions, mrope_position_delta = (
+                        MRotaryEmbedding.get_input_positions(
+                            input_tokens=self.input_ids[
+                                extend_start_loc : extend_start_loc + extend_seq_len
+                            ],
+                            image_grid_thw=image_inputs.image_grid_thws,
+                            vision_start_token_id=hf_config.vision_start_token_id,
+                            spatial_merge_size=hf_config.vision_config.spatial_merge_size,
+                            context_len=0,
+                        )
+                    )
+                    batch.image_inputs[i].mrope_position_delta = mrope_position_delta
+                mrope_positions_list[i] = mrope_positions
+        self.mrope_positions = torch.concat(
+            [torch.tensor(pos, device=device) for pos in mrope_positions_list],
+            axis=1,
+        )
+        self.mrope_positions = self.mrope_positions.to(torch.int64)
+
 
 def compute_position_triton(
     extend_prefix_lens: torch.Tensor, extend_seq_lens: torch.Tensor, extend_seq_lens_sum

python/sglang/srt/model_executor/model_runner.py

@@ -52,6 +52,7 @@ from sglang.srt.mem_cache.memory_pool import (
     MLATokenToKVPool,
     ReqToTokenPool,
 )
+from sglang.srt.model_executor.cuda_graph_runner import CudaGraphRunner
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader import get_model
 from sglang.srt.server_args import ServerArgs
@@ -714,8 +715,6 @@ class ModelRunner:
 
     def init_cuda_graphs(self):
         """Capture cuda graphs."""
-        from sglang.srt.model_executor.cuda_graph_runner import CudaGraphRunner
-
         self.cuda_graph_runner = None
 
         if not self.is_generation:

python/sglang/srt/speculative/build_eagle_tree.py

@@ -79,11 +79,13 @@ __global__ void build_tree(Tensor<long, 2> parent_list, Tensor<long, 2> selected
 )
 
 
-def build_tree_kernel(parent_list, top_score_index, seq_lens, topk, depth, draft_token):
+def build_tree_kernel(
+    parent_list, top_score_index, seq_lens, seq_lens_sum, topk, depth, draft_token
+):
     bs = seq_lens.numel()
     device = parent_list.device
     tree_mask = torch.full(
-        (torch.sum(seq_lens).item() * draft_token + draft_token * draft_token * bs,),
+        (seq_lens_sum * draft_token + draft_token * draft_token * bs,),
         True,
         device=device,
     )

python/sglang/srt/speculative/eagle_draft_cuda_graph_runner.py

+from __future__ import annotations
+
+import bisect
+import time
+from typing import TYPE_CHECKING, Callable
+
+import torch
+
+from sglang.srt.model_executor.cuda_graph_runner import (
+    CudaGraphRunner,
+    get_batch_sizes_to_capture,
+    get_global_graph_memory_pool,
+    set_global_graph_memory_pool,
+    set_torch_compile_config,
+)
+from sglang.srt.model_executor.forward_batch_info import (
+    CaptureHiddenMode,
+    ForwardBatch,
+    ForwardMode,
+)
+from sglang.srt.speculative.eagle_utils import EagleDraftInput
+
+if TYPE_CHECKING:
+    from sglang.srt.model_executor.model_runner import ModelRunner
+    from sglang.srt.speculative.eagle_worker import EAGLEWorker
+
+
+class EAGLEDraftCudaGraphRunner:
+    def __init__(self, eagle_worker: EAGLEWorker):
+        # Parse args
+        self.eagle_worker = eagle_worker
+        self.model_runner = model_runner = eagle_worker.model_runner
+        self.graphs = {}
+        self.output_buffers = {}
+        self.enable_torch_compile = model_runner.server_args.enable_torch_compile
+        self.disable_padding = model_runner.server_args.disable_cuda_graph_padding
+        self.tp_size = self.model_runner.tp_size
+        self.dp_size = model_runner.server_args.dp_size
+        self.topk = model_runner.server_args.speculative_eagle_topk
+        self.speculative_num_steps = model_runner.server_args.speculative_num_steps
+        server_args = model_runner.server_args
+
+        assert self.disable_padding
+
+        # Batch sizes to capture
+        self.capture_bs, self.compile_bs = get_batch_sizes_to_capture(model_runner)
+        self.num_tokens_per_bs = server_args.speculative_eagle_topk
+
+        # Attention backend
+        self.max_bs = max(self.capture_bs)
+        self.max_num_token = self.max_bs * self.num_tokens_per_bs
+        self.model_runner.draft_attn_backend.init_cuda_graph_state(self.max_num_token)
+        self.seq_len_fill_value = self.model_runner.draft_attn_backend.attn_backends[
+            0
+        ].get_cuda_graph_seq_len_fill_value()
+
+        if self.enable_torch_compile:
+            set_torch_compile_config()
+
+        # Graph inputs
+        with torch.device("cuda"):
+            self.input_ids = torch.zeros((self.max_num_token,), dtype=torch.int64)
+            self.req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int32)
+            self.seq_lens = torch.full(
+                (self.max_bs,), self.seq_len_fill_value, dtype=torch.int32
+            )
+            self.out_cache_loc = torch.zeros(
+                (self.max_num_token * self.speculative_num_steps,), dtype=torch.int64
+            )
+            self.positions = torch.zeros((self.max_num_token,), dtype=torch.int64)
+            self.topk_p = torch.zeros((self.max_bs, self.topk), dtype=torch.float32)
+            self.topk_index = torch.zeros((self.max_bs, self.topk), dtype=torch.int64)
+            self.hidden_states = torch.zeros(
+                (self.max_bs, self.model_runner.model_config.hidden_size),
+                dtype=self.model_runner.dtype,
+            )
+
+        # Capture
+        try:
+            self.capture()
+        except RuntimeError as e:
+            raise Exception(
+                f"Capture cuda graph failed: {e}\n"
+                "Possible solutions:\n"
+                "1. disable cuda graph by --disable-cuda-graph\n"
+                "2. set --mem-fraction-static to a smaller value (e.g., 0.8 or 0.7)\n"
+                "3. disable torch compile by not using --enable-torch-compile\n"
+                "Open an issue on GitHub https://github.com/sgl-project/sglang/issues/new/choose \n"
+            )
+
+    def can_run(self, forward_batch: ForwardBatch):
+        is_bs_supported = (
+            forward_batch.batch_size in self.graphs
+            if self.disable_padding
+            else forward_batch.batch_size <= self.max_bs
+        )
+        return is_bs_supported
+
+    def capture(self):
+        CudaGraphRunner.capture(self)
+
+    def capture_one_batch_size(self, num_seqs: int, forward: Callable):
+        graph = torch.cuda.CUDAGraph()
+        stream = self.stream
+        num_tokens = num_seqs * self.num_tokens_per_bs
+
+        # Graph inputs
+        req_pool_indices = self.req_pool_indices[:num_seqs]
+        seq_lens = self.seq_lens[:num_seqs]
+        out_cache_loc = self.out_cache_loc[: num_tokens * self.speculative_num_steps]
+        positions = self.positions[:num_tokens]
+        topk_p = self.topk_p[:num_seqs]
+        topk_index = self.topk_index[:num_seqs]
+        hidden_states = self.hidden_states[:num_seqs]
+
+        spec_info = EagleDraftInput(
+            topk_p=topk_p,
+            topk_index=topk_index,
+            hidden_states=hidden_states,
+        )
+
+        # Forward batch
+        forward_batch = ForwardBatch(
+            forward_mode=ForwardMode.DECODE,
+            batch_size=num_seqs,
+            input_ids=None,
+            req_pool_indices=req_pool_indices,
+            seq_lens=seq_lens,
+            req_to_token_pool=self.model_runner.req_to_token_pool,
+            token_to_kv_pool=self.model_runner.token_to_kv_pool,
+            out_cache_loc=out_cache_loc,
+            seq_lens_sum=seq_lens.sum(),
+            return_logprob=False,
+            positions=positions,
+            spec_algorithm=self.model_runner.spec_algorithm,
+            spec_info=spec_info,
+            capture_hidden_mode=(
+                spec_info.capture_hidden_mode if spec_info else CaptureHiddenMode.NULL
+            ),
+        )
+
+        # Attention backend
+        self.model_runner.draft_attn_backend.init_forward_metadata_capture_cuda_graph(
+            forward_batch
+        )
+
+        # Run and capture
+        def run_once():
+            # Backup two fileds, which will be modified in-place in `draft_forward`.
+            output_cache_loc_backup = forward_batch.out_cache_loc
+            hidden_states_backup = forward_batch.spec_info.hidden_states
+
+            ret = self.eagle_worker.draft_forward(forward_batch)
+
+            forward_batch.out_cache_loc = output_cache_loc_backup
+            forward_batch.spec_info.hidden_states = hidden_states_backup
+            return ret
+
+        for _ in range(2):
+            torch.cuda.synchronize()
+            self.model_runner.tp_group.barrier()
+
+            run_once()
+
+            torch.cuda.synchronize()
+            self.model_runner.tp_group.barrier()
+
+        torch.cuda.synchronize()
+        self.model_runner.tp_group.barrier()
+
+        with torch.cuda.graph(
+            graph, pool=get_global_graph_memory_pool(), stream=stream
+        ):
+            out = run_once()
+
+        torch.cuda.synchronize()
+        self.model_runner.tp_group.barrier()
+
+        set_global_graph_memory_pool(graph.pool())
+        return graph, out
+
+    def replay(self, forward_batch: ForwardBatch):
+        assert forward_batch.out_cache_loc is not None
+        raw_bs = forward_batch.batch_size
+        raw_num_token = raw_bs * self.num_tokens_per_bs
+
+        # Pad
+        index = bisect.bisect_left(self.capture_bs, raw_bs)
+        bs = self.capture_bs[index]
+        if bs != raw_bs:
+            self.seq_lens.fill_(1)
+            self.out_cache_loc.zero_()
+
+        # Common inputs
+        self.req_pool_indices[:raw_bs].copy_(forward_batch.req_pool_indices)
+        self.seq_lens[:raw_bs].copy_(forward_batch.seq_lens)
+        self.out_cache_loc[: raw_num_token * self.speculative_num_steps].copy_(
+            forward_batch.out_cache_loc
+        )
+        self.positions[:raw_num_token].copy_(forward_batch.positions)
+        self.topk_p[:raw_bs].copy_(forward_batch.spec_info.topk_p)
+        self.topk_index[:raw_bs].copy_(forward_batch.spec_info.topk_index)
+        self.hidden_states[:raw_bs].copy_(forward_batch.spec_info.hidden_states)
+
+        # Attention backend
+        self.model_runner.draft_attn_backend.init_forward_metadata_replay_cuda_graph(
+            forward_batch
+        )
+
+        # Replay
+        self.graphs[bs].replay()
+
+        return self.output_buffers[bs]

python/sglang/srt/speculative/eagle_utils.py

 from __future__ import annotations
 
+import dataclasses
 from typing import TYPE_CHECKING, List
 
 import torch
@@ -9,201 +10,33 @@ import triton.language as tl
 from sglang.srt.layers.attention.flashinfer_backend import (
     create_flashinfer_kv_indices_triton,
 )
-from sglang.srt.model_executor.forward_batch_info import ForwardMode
+from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode
 from sglang.srt.speculative.build_eagle_tree import build_tree_kernel
-from sglang.srt.speculative.spec_info import SpecInfo
 
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import ScheduleBatch
-    from sglang.srt.server_args import ServerArgs
 
 
-@triton.jit
-def eagle_verify_retrive(
-    retrive_index,
-    accept_mask,
-    retrive_cum_len,
-    accept_index,
-    accept_length,
-    extract_index,
-    max_len: tl.constexpr,
-    draft_token_num: tl.constexpr,
-    max_len_upper: tl.constexpr,
-):
-    pid = tl.program_id(axis=0)
-
-    retrive_end = tl.load(retrive_cum_len + pid + 1)
-    retrive_start = tl.load(retrive_cum_len + pid)
-    retrive_len = retrive_end - retrive_start
-    accept_ptr = accept_mask + retrive_start
-    accept_offset = tl.arange(0, draft_token_num)
-    accept_load_mask = accept_offset < retrive_len
-    accept_len_list = tl.load(
-        accept_ptr + accept_offset, mask=accept_load_mask, other=-1
-    )
-
-    accept_len = tl.max(accept_len_list)
-    max_index = tl.argmax(accept_len_list, axis=0, tie_break_left=True)
-    # triton is not support argmax with tie_break_right, so I need implement it by some way
-    mask_max = accept_len_list == accept_len
-
-    count_mask = tl.full(shape=[draft_token_num], value=0, dtype=tl.int32)
-    count = tl.sum(tl.where(mask_max, 1, count_mask))
-    if count > 1:
-        index = tl.arange(0, draft_token_num)
-        mask_left = index != max_index
-        remained_index = tl.where(mask_max and mask_left, index, 0)
-        max_index = tl.max(remained_index)
-
-    tl.store(accept_length + pid, accept_len)
-    retrive_index_ptr = retrive_index + (retrive_start + max_index) * max_len
-    retrive_offset = tl.arange(0, max_len_upper)
-    retrive_load_mask = retrive_offset < accept_len + 1
-    data = tl.load(retrive_index_ptr + retrive_offset, mask=retrive_load_mask)
-
-    tl.store(
-        accept_index + pid * max_len + retrive_offset, data, mask=retrive_load_mask
-    )
-
-    extract_load_ptr = accept_index + pid * max_len + accept_len
-    if accept_len == max_len - 1:
-        extract_data = tl.load(extract_load_ptr - 1)
-        tl.store(extract_index + pid * 2, extract_data)
-        extract_data = tl.load(extract_load_ptr)
-        tl.store(extract_index + pid * 2 + 1, extract_data)
-
-    else:
-        extract_data = tl.load(extract_load_ptr)
-        tl.store(extract_index + pid * 2, extract_data)
-
-
-@triton.jit
-def create_extend_spec_info(
-    verified_id,
-    seq_len,
-    accept_len,
-    accept_len_cum,
-    positions,
-    new_verified_id,
-    accept_len_upper: tl.constexpr,
-):
-    pid = tl.program_id(axis=0)
-    offset = 0 if pid == 0 else tl.load(accept_len_cum + pid - 1)
-    seq_length = tl.load(seq_len + pid)
-    accept_length = tl.load(accept_len + pid)
-    positions_ptr = positions + offset
-    data = tl.arange(0, accept_len_upper)
-    mask = data < accept_length
-    tl.store(positions_ptr + data, seq_length - accept_length + data, mask)
-
-    offset = tl.load(accept_len_cum + pid) - 1
-    verified_id_data = tl.load(verified_id + offset)
-    tl.store(new_verified_id + pid, verified_id_data)
-
-
-@triton.jit
-def assign_req_to_token_pool(
-    req_pool_indices,
-    req_to_token,
-    start_offset,
-    end_offset,
-    out_cache_loc,
-    pool_len: tl.constexpr,
-    bs_upper: tl.constexpr,
-):
-    BLOCK_SIZE: tl.constexpr = 32
-    pid = tl.program_id(axis=0)
-    kv_start = tl.load(start_offset + pid)
-    kv_end = tl.load(end_offset + pid)
-    token_pool = req_to_token + tl.load(req_pool_indices + pid) * pool_len
-
-    length_offset = tl.arange(0, bs_upper)
-    start = tl.load(start_offset + length_offset, mask=length_offset < pid)
-    end = tl.load(end_offset + length_offset, mask=length_offset < pid)
-    out_offset = tl.sum(end - start, axis=0)
-
-    out_cache_ptr = out_cache_loc + out_offset
-
-    save_offset = tl.arange(0, BLOCK_SIZE) + kv_start
-    load_offset = tl.arange(0, BLOCK_SIZE)
-
-    num_loop = tl.cdiv(kv_end - kv_start, BLOCK_SIZE)
-    for _ in range(num_loop):
-        mask = save_offset < kv_end
-        data = tl.load(out_cache_ptr + load_offset, mask=mask)
-        tl.store(token_pool + save_offset, data, mask=mask)
-        save_offset += BLOCK_SIZE
-        load_offset += BLOCK_SIZE
-
-
-@triton.jit
-def generate_draft_decode_kv_indices(
-    req_pool_indices,
-    req_to_token,
-    paged_kernel_lens,
-    kv_indices,
-    iters: tl.constexpr,
-    topk: tl.constexpr,
-    pool_len: tl.constexpr,
-    bs_upper: tl.constexpr,
-    iter_upper: tl.constexpr,
-):
-    BLOCK_SIZE: tl.constexpr = 128
-    bid = tl.program_id(axis=0)
-    topk_id = tl.program_id(axis=1)
-
-    load_offset = tl.arange(0, bs_upper)
-    seq_lens = tl.load(paged_kernel_lens + load_offset, mask=load_offset < bid)
-    seq_len = tl.load(paged_kernel_lens + bid)
-    cum_seq_len = tl.sum(seq_lens)
-
-    kv_offset = cum_seq_len * topk + bid * iters * topk + topk_id * (seq_len + iters)
-    kv_ptr = kv_indices + kv_offset
-    token_pool_ptr = req_to_token + tl.load(req_pool_indices + bid) * pool_len
-
-    kv_offset = tl.arange(0, BLOCK_SIZE)
-    num_loop = tl.cdiv(seq_len, BLOCK_SIZE)
-    for _ in range(num_loop):
-        mask = kv_offset < seq_len
-        data = tl.load(token_pool_ptr + kv_offset, mask=mask)
-        tl.store(kv_ptr + kv_offset, data, mask=mask)
-        kv_offset += BLOCK_SIZE
-
-    extend_offset = tl.arange(0, iter_upper)
-    extend_data = tl.load(
-        token_pool_ptr + seq_len + tl.arange(0, iter_upper) * topk + topk_id,
-        mask=extend_offset < iters,
-    )
-    tl.store(kv_ptr + seq_len + extend_offset, extend_data, mask=extend_offset < iters)
-
-
-class EAGLEDraftInput(SpecInfo):
-    def __init__(self):
-        self.prev_mode = ForwardMode.DECODE
-
-        self.scores: torch.Tensor = None
-        self.score_list: List[torch.Tensor] = []
-        self.token_list: List[torch.Tensor] = []
-        self.origin_score_list: List[torch.Tensor] = []  # used for sampling
-        self.parents_list: List[torch.Tensor] = []
-        self.cache_list: List[torch.Tenor] = []
-        self.iter = 0
-
+@dataclasses.dataclass
+class EagleDraftInput:
+    # The inputs for decode
+    # shape: (b, topk)
+    topk_p: torch.Tensor = None
+    topk_index: torch.Tensor = None
     # shape: (b, hidden_size)
-        self.hidden_states: torch.Tensor = None
-        # shape: (b,)
-        self.verified_id: torch.Tensor = None
-        # shape: (b, vocab_size)
-        self.sample_output: torch.Tensor = None
+    hidden_states: torch.Tensor = None
+    capture_hidden_mode: CaptureHiddenMode = CaptureHiddenMode.FULL
 
-        self.positions: torch.Tensor = None
-        self.accept_length: torch.Tensor = None
-        self.accept_length_cpu: List[int] = None
+    # Inputs for extend
+    # shape: (b,)
+    verified_id: torch.Tensor = None
+    accept_length: torch.Tensor = None
+    accept_length_cpu: List[int] = None
 
-    def load_server_args(self, server_args: ServerArgs):
-        self.topk: int = server_args.speculative_eagle_topk
-        self.num_verify_token: int = server_args.speculative_num_draft_tokens
-        self.spec_steps = server_args.speculative_num_steps
+    # Inputs for the attention backends
+    # shape: (b + 1,)
+    kv_indptr: torch.Tensor = None
+    kv_indices: torch.Tensor = None
 
     def prepare_for_extend(self, batch: ScheduleBatch):
         req_pool_indices = batch.alloc_req_slots(len(batch.reqs))
@@ -231,91 +64,7 @@ class EAGLEDraftInput(SpecInfo):
         assert len(batch.extend_lens) == 1
         batch.input_ids = torch.concat((batch.input_ids[1:], self.verified_id))
 
-    def filter_batch(
-        self,
-        new_indices: torch.Tensor,
-    ):
-        self.sample_output = self.sample_output[: len(new_indices)]
-        self.hidden_states = self.hidden_states[: len(new_indices)]
-        self.verified_id = self.verified_id[: len(new_indices)]
-
-    def prepare_for_decode(self, batch: ScheduleBatch):
-        prob = self.sample_output  # shape: (b * top_k, vocab) or (b, vocab)
-        top = torch.topk(prob, self.topk, dim=-1)
-        topk_index, topk_p = (
-            top.indices,
-            top.values,
-        )  # shape: (b * top_k, top_k) or (b, top_k)
-
-        if self.prev_mode.is_decode():
-            scores = torch.mul(
-                self.scores.unsqueeze(2), topk_p.reshape(-1, self.topk, self.topk)
-            )  # (b, topk, 1) x (b, topk ,topk) -> (b, topk, topk)
-            topk_cs = torch.topk(
-                scores.flatten(start_dim=1), self.topk, dim=-1
-            )  # (b, topk)
-            topk_cs_index, topk_cs_p = topk_cs.indices, topk_cs.values
-
-            selected_input_index = topk_cs_index.flatten() // self.topk + torch.arange(
-                0, batch.batch_size() * self.topk, step=self.topk, device="cuda"
-            ).repeat_interleave(self.topk)
-
-            batch.spec_info.hidden_states = batch.spec_info.hidden_states[
-                selected_input_index, :
-            ]
-
-            topk_index = topk_index.reshape(-1, self.topk**2)
-            batch.input_ids = torch.gather(
-                topk_index, index=topk_cs_index, dim=1
-            ).flatten()
-            batch.out_cache_loc = batch.alloc_token_slots(len(batch.input_ids))
-
-            self.scores = topk_cs_p
-            self.score_list.append(scores)  # (b, topk, topk)
-            self.token_list.append(topk_index)  # (b, topk * topk)
-            self.origin_score_list.append(topk_p.reshape(topk_index.shape))
-            self.parents_list.append(
-                topk_cs_index + (self.topk**2 * (self.iter - 1) + self.topk)
-            )  # shape: (b, topk)
-        else:
-            # ForwardMode.EXTEND or ForwardMode.DRAFT_EXTEND
-            batch.spec_info.hidden_states = (
-                batch.spec_info.hidden_states.repeat_interleave(self.topk, dim=0)
-            )
-
-            batch.input_ids = topk_index.flatten()
-            batch.out_cache_loc = batch.alloc_token_slots(topk_index.numel())
-
-            self.scores = topk_p  # shape: (b, topk)
-            self.score_list.append(topk_p.unsqueeze(1))  # shape: (b, 1, topk)
-            self.token_list.append(topk_index)  # shape: (b, topk)
-            self.origin_score_list.append(topk_p)
-            self.parents_list.append(
-                torch.arange(-1, self.topk, dtype=torch.long, device="cuda")
-                .unsqueeze(0)
-                .repeat(self.scores.shape[0], 1)
-            )  # shape: (b, topk + 1)
-        self.cache_list.append(batch.out_cache_loc)
-        self.positions = (
-            batch.seq_lens[:, None]
-            + torch.full(
-                [1, self.topk], fill_value=self.iter, device="cuda", dtype=torch.long
-            )
-        ).flatten()
-
-        bs = len(batch.seq_lens)
-        assign_req_to_token_pool[(bs,)](
-            batch.req_pool_indices,
-            batch.req_to_token_pool.req_to_token,
-            batch.seq_lens + self.topk * self.iter,
-            batch.seq_lens + self.topk * (self.iter + 1),
-            batch.out_cache_loc,
-            batch.req_to_token_pool.req_to_token.shape[1],
-            triton.next_power_of_2(bs),
-        )
-        self.iter += 1
-
-    def prepare_extend_after_decode(self, batch: ScheduleBatch):
+    def prepare_extend_after_decode(self, batch: ScheduleBatch, speculative_num_steps):
         batch.out_cache_loc = batch.alloc_token_slots(self.verified_id.numel())
         accept_length_cpu = batch.spec_info.accept_length_cpu
         batch.extend_lens = [x + 1 for x in accept_length_cpu]
@@ -348,86 +97,13 @@ class EAGLEDraftInput(SpecInfo):
             torch.cumsum(self.accept_length, axis=0, dtype=torch.int),
             self.positions,
             new_verified_id,
-            triton.next_power_of_2(self.spec_steps + 1),
+            triton.next_power_of_2(speculative_num_steps + 1),
         )
 
         batch.seq_lens_sum = sum(seq_lens_cpu)
         batch.input_ids = self.verified_id
         self.verified_id = new_verified_id
 
-    def prepare_for_verify(self, batch: ScheduleBatch):
-        score_list = torch.cat(self.score_list, dim=1).flatten(
-            1
-        )  # b, n, topk; n= 1+(self.iter-1)*self.topk
-        ss_token_list = torch.cat(
-            self.token_list, dim=1
-        )  # b, (self.topk+(self.iter-1)*self.topk)
-        origin_token_list = torch.cat(self.origin_score_list, dim=1)
-        top_scores = torch.topk(score_list, self.num_verify_token - 1, dim=-1)
-        top_scores_index = top_scores.indices
-        top_scores_index = torch.sort(top_scores_index).values
-
-        draft_tokens = torch.gather(ss_token_list, index=top_scores_index, dim=1)
-        scores = torch.gather(origin_token_list, index=top_scores_index, dim=1)
-        draft_tokens = torch.cat((self.verified_id.unsqueeze(1), draft_tokens), dim=1)
-        parent_list = torch.cat(self.parents_list[:-1], dim=1)
-
-        tree_mask, position, retrive_index, retrive_cum_len = build_tree_kernel(
-            parent_list,
-            top_scores_index,
-            batch.seq_lens,
-            self.topk,
-            self.iter - 1,
-            self.num_verify_token,
-        )
-
-        return EagleVerifyInput(
-            draft_tokens.flatten(),
-            scores.flatten(),
-            tree_mask,
-            position,
-            retrive_index,
-            retrive_cum_len,
-            self.num_verify_token,
-        )
-
-    def generate_attn_arg_decode(
-        self,
-        req_pool_indices: torch.Tensor,
-        paged_kernel_lens: torch.Tensor,
-        req_to_token: torch.Tensor,
-    ):
-        seq_num = req_pool_indices.numel()
-        bs = self.topk * req_pool_indices.numel()
-        seq_len = self.positions.reshape(-1).contiguous()
-
-        cum_kv_seq_len = torch.zeros((bs + 1,), dtype=torch.int32, device="cuda")
-        cum_kv_seq_len[1:] = torch.cumsum(seq_len + 1, dim=0)
-        total_len = torch.sum(paged_kernel_lens).item()
-
-        kv_indices = torch.empty(
-            (total_len * self.topk + seq_num * self.iter * self.topk,),
-            dtype=torch.int32,
-            device="cuda",
-        )
-
-        generate_draft_decode_kv_indices[(req_pool_indices.numel(), self.topk)](
-            req_pool_indices,
-            req_to_token,
-            paged_kernel_lens,
-            kv_indices,
-            self.iter,
-            self.topk,
-            req_to_token.shape[1],
-            triton.next_power_of_2(seq_num),
-            triton.next_power_of_2(self.spec_steps),
-        )
-        return bs, kv_indices, cum_kv_seq_len
-
-    def clear_draft_cache(self, batch):
-        draft_cache = torch.cat(self.cache_list, dim=0)
-        batch.token_to_kv_pool.free(draft_cache)
-
     def generate_attn_arg_prefill(
         self,
         req_pool_indices: torch.Tensor,
@@ -454,12 +130,18 @@ class EAGLEDraftInput(SpecInfo):
 
         return kv_indices, cum_kv_seq_len, qo_indptr, None
 
-    def merge_batch(self, spec_info: EAGLEDraftInput):
+    def filter_batch(self, new_indices: torch.Tensor):
+        self.topk_p = self.topk_p[: len(new_indices)]
+        self.topk_index = self.topk_index[: len(new_indices)]
+        self.hidden_states = self.hidden_states[: len(new_indices)]
+        self.verified_id = self.verified_id[: len(new_indices)]
+
+    def merge_batch(self, spec_info: EagleDraftInput):
         if self.hidden_states is None:
             self.hidden_states = spec_info.hidden_states
             self.verified_id = spec_info.verified_id
-            self.sample_output = spec_info.sample_output
-            self.prev_mode = spec_info.prev_mode
+            self.topk_p = spec_info.topk_p
+            self.topk_index = spec_info.topk_index
             return
         if spec_info.hidden_states is None:
             return
@@ -467,32 +149,68 @@ class EAGLEDraftInput(SpecInfo):
             [self.hidden_states, spec_info.hidden_states], axis=0
         )
         self.verified_id = torch.cat([self.verified_id, spec_info.verified_id], axis=0)
-        self.sample_output = torch.cat([self.sample_output, spec_info.sample_output])
-
-
-class EagleVerifyInput(SpecInfo):
-    def __init__(
-        self,
-        draft_token: torch.Tensor,
-        draft_score: torch.Tensor,
-        tree_mask: torch.Tensor,
-        positions: torch.Tensor,
-        retrive_index: torch.Tensor,
-        retrive_cum_len: torch.Tensor,
-        draft_token_num: int,
+        self.topk_p = torch.cat([self.topk_p, spec_info.topk_p])
+        self.topk_index = torch.cat([self.topk_index, spec_info.topk_index])
+
+
+@dataclasses.dataclass
+class EagleVerifyInput:
+    draft_token: torch.Tensor
+    custom_mask: torch.Tensor
+    positions: torch.Tensor
+    retrive_index: torch.Tensor
+    retrive_cum_len: torch.Tensor
+    draft_token_num: int
+    capture_hidden_mode: CaptureHiddenMode
+
+    @classmethod
+    def create(
+        cls,
+        verified_id: torch.Tensor,
+        score_list: List[torch.Tensor],
+        token_list: List[torch.Tensor],
+        parents_list: List[torch.Tensor],
+        seq_lens: torch.Tensor,
+        seq_lens_sum: int,
+        topk: int,
+        spec_steps: int,
+        num_verify_token: int,
     ):
-        self.draft_token = draft_token
-        self.draft_score = draft_score
-        self.custom_mask = tree_mask
-        self.positions = positions
-        self.retrive_index = retrive_index
-        self.retrive_cum_len = retrive_cum_len
-        self.draft_token_num = draft_token_num
+        score_list = torch.cat(score_list, dim=1).flatten(
+            1
+        )  # b, n, topk; n= 1 + (num_steps-1) * self.topk
+        ss_token_list = torch.cat(
+            token_list, dim=1
+        )  # b, (self.topk + (num_steps-1) * self.topk)
+        top_scores = torch.topk(score_list, num_verify_token - 1, dim=-1)
+        top_scores_index = top_scores.indices
+        top_scores_index = torch.sort(top_scores_index).values
+        draft_tokens = torch.gather(ss_token_list, index=top_scores_index, dim=1)
+        draft_tokens = torch.cat((verified_id.unsqueeze(1), draft_tokens), dim=1)
+        parent_list = torch.cat(parents_list[:-1], dim=1)
+        tree_mask, position, retrive_index, retrive_cum_len = build_tree_kernel(
+            parent_list,
+            top_scores_index,
+            seq_lens,
+            seq_lens_sum,
+            topk,
+            spec_steps,
+            num_verify_token,
+        )
+        return cls(
+            draft_tokens.flatten(),
+            tree_mask,
+            position,
+            retrive_index,
+            retrive_cum_len,
+            num_verify_token,
+            CaptureHiddenMode.FULL,
+        )
 
     def prepare_for_verify(self, batch: ScheduleBatch):
         batch.input_ids = self.draft_token
         batch.out_cache_loc = batch.alloc_token_slots(batch.input_ids.numel())
-        bs = batch.seq_lens.numel()
+        bs = batch.batch_size()
         assign_req_to_token_pool[(bs,)](
             batch.req_pool_indices,
             batch.req_to_token_pool.req_to_token,
@@ -573,7 +291,6 @@ class EagleVerifyInput(SpecInfo):
             triton.next_power_of_2(max_draft_len),
         )
 
-        draft_input = EAGLEDraftInput()
         new_accept_index = []
         unfinished_index = []
         finished_extend_len = {}  # {rid:accept_length + 1}
@@ -625,10 +342,11 @@ class EagleVerifyInput(SpecInfo):
         )
         batch.seq_lens.add_(accept_length + 1)
 
+        draft_input = EagleDraftInput()
         if len(new_accept_index) > 0:
             new_accept_index = torch.tensor(new_accept_index, device="cuda")
-            draft_input.verified_id = predict[new_accept_index]
             draft_input.hidden_states = batch.spec_info.hidden_states[new_accept_index]
+            draft_input.verified_id = predict[new_accept_index]
             draft_input.accept_length = accept_length[unfinished_index]
             draft_input.accept_length_cpu = [
                 accept_length_cpu[i] for i in unfinished_index
@@ -646,3 +364,269 @@ class EagleVerifyInput(SpecInfo):
             finished_extend_len,
             accept_length_cpu,
         )
+
+
+@triton.jit
+def eagle_verify_retrive(
+    retrive_index,
+    accept_mask,
+    retrive_cum_len,
+    accept_index,
+    accept_length,
+    extract_index,
+    max_len: tl.constexpr,
+    draft_token_num: tl.constexpr,
+    max_len_upper: tl.constexpr,
+):
+    pid = tl.program_id(axis=0)
+
+    retrive_end = tl.load(retrive_cum_len + pid + 1)
+    retrive_start = tl.load(retrive_cum_len + pid)
+    retrive_len = retrive_end - retrive_start
+    accept_ptr = accept_mask + retrive_start
+    accept_offset = tl.arange(0, draft_token_num)
+    accept_load_mask = accept_offset < retrive_len
+    accept_len_list = tl.load(
+        accept_ptr + accept_offset, mask=accept_load_mask, other=-1
+    )
+
+    accept_len = tl.max(accept_len_list)
+    max_index = tl.argmax(accept_len_list, axis=0, tie_break_left=True)
+    # triton is not support argmax with tie_break_right, so I need implement it by some way
+    mask_max = accept_len_list == accept_len
+
+    count_mask = tl.full(shape=[draft_token_num], value=0, dtype=tl.int32)
+    count = tl.sum(tl.where(mask_max, 1, count_mask))
+    if count > 1:
+        index = tl.arange(0, draft_token_num)
+        mask_left = index != max_index
+        remained_index = tl.where(mask_max and mask_left, index, 0)
+        max_index = tl.max(remained_index)
+
+    tl.store(accept_length + pid, accept_len)
+    retrive_index_ptr = retrive_index + (retrive_start + max_index) * max_len
+    retrive_offset = tl.arange(0, max_len_upper)
+    retrive_load_mask = retrive_offset < accept_len + 1
+    data = tl.load(retrive_index_ptr + retrive_offset, mask=retrive_load_mask)
+
+    tl.store(
+        accept_index + pid * max_len + retrive_offset, data, mask=retrive_load_mask
+    )
+
+    extract_load_ptr = accept_index + pid * max_len + accept_len
+    if accept_len == max_len - 1:
+        extract_data = tl.load(extract_load_ptr - 1)
+        tl.store(extract_index + pid * 2, extract_data)
+        extract_data = tl.load(extract_load_ptr)
+        tl.store(extract_index + pid * 2 + 1, extract_data)
+
+    else:
+        extract_data = tl.load(extract_load_ptr)
+        tl.store(extract_index + pid * 2, extract_data)
+
+
+@triton.jit
+def create_extend_spec_info(
+    verified_id,
+    seq_len,
+    accept_len,
+    accept_len_cum,
+    positions,
+    new_verified_id,
+    accept_len_upper: tl.constexpr,
+):
+    pid = tl.program_id(axis=0)
+    offset = 0 if pid == 0 else tl.load(accept_len_cum + pid - 1)
+    seq_length = tl.load(seq_len + pid)
+    accept_length = tl.load(accept_len + pid)
+    positions_ptr = positions + offset
+    data = tl.arange(0, accept_len_upper)
+    mask = data < accept_length
+    tl.store(positions_ptr + data, seq_length - accept_length + data, mask)
+
+    offset = tl.load(accept_len_cum + pid) - 1
+    verified_id_data = tl.load(verified_id + offset)
+    tl.store(new_verified_id + pid, verified_id_data)
+
+
+@triton.jit
+def assign_req_to_token_pool(
+    req_pool_indices,
+    req_to_token,
+    start_offset,
+    end_offset,
+    out_cache_loc,
+    pool_len: tl.constexpr,
+    bs_upper: tl.constexpr,
+):
+    BLOCK_SIZE: tl.constexpr = 32
+    pid = tl.program_id(axis=0)
+    kv_start = tl.load(start_offset + pid)
+    kv_end = tl.load(end_offset + pid)
+    token_pool = req_to_token + tl.load(req_pool_indices + pid) * pool_len
+
+    length_offset = tl.arange(0, bs_upper)
+    start = tl.load(start_offset + length_offset, mask=length_offset < pid)
+    end = tl.load(end_offset + length_offset, mask=length_offset < pid)
+    out_offset = tl.sum(end - start, axis=0)
+
+    out_cache_ptr = out_cache_loc + out_offset
+
+    save_offset = tl.arange(0, BLOCK_SIZE) + kv_start
+    load_offset = tl.arange(0, BLOCK_SIZE)
+
+    num_loop = tl.cdiv(kv_end - kv_start, BLOCK_SIZE)
+    for _ in range(num_loop):
+        mask = save_offset < kv_end
+        data = tl.load(out_cache_ptr + load_offset, mask=mask)
+        tl.store(token_pool + save_offset, data, mask=mask)
+        save_offset += BLOCK_SIZE
+        load_offset += BLOCK_SIZE
+
+
+@triton.jit
+def assign_draft_cache_locs(
+    req_pool_indices,
+    req_to_token,
+    seq_lens,
+    out_cache_loc,
+    pool_len: tl.constexpr,
+    topk: tl.constexpr,
+    speculative_num_steps: tl.constexpr,
+):
+    BLOCK_SIZE: tl.constexpr = 32
+    pid = tl.program_id(axis=0)
+    kv_start = tl.load(seq_lens + pid)
+    kv_end = tl.load(seq_lens + pid) + topk * speculative_num_steps
+    token_pool = req_to_token + tl.load(req_pool_indices + pid) * pool_len
+    out_cache_ptr = out_cache_loc + pid * topk * speculative_num_steps
+
+    num_loop = tl.cdiv(topk * speculative_num_steps, BLOCK_SIZE)
+    for i in range(num_loop):
+        save_offset = tl.arange(0, BLOCK_SIZE) + i * BLOCK_SIZE + kv_start
+        load_offset = tl.arange(0, BLOCK_SIZE) + i * BLOCK_SIZE
+        mask = save_offset < kv_end
+        data = tl.load(out_cache_ptr + load_offset, mask=mask)
+        tl.store(token_pool + save_offset, data, mask=mask)
+
+
+@triton.jit
+def generate_draft_decode_kv_indices(
+    req_pool_indices,
+    req_to_token,
+    paged_kernel_lens,
+    kv_indices,
+    kv_indptr,
+    positions,
+    num_seqs: tl.constexpr,
+    topk: tl.constexpr,
+    pool_len: tl.constexpr,
+    kv_indices_stride: tl.constexpr,
+    kv_indptr_stride: tl.constexpr,
+    bs_upper: tl.constexpr,
+    iter_upper: tl.constexpr,
+    num_tokens_upper: tl.constexpr,
+):
+    BLOCK_SIZE: tl.constexpr = 128
+    iters = tl.program_id(axis=0)
+    bid = tl.program_id(axis=1)
+    topk_id = tl.program_id(axis=2)
+
+    kv_indices += kv_indices_stride * iters
+    kv_indptr += kv_indptr_stride * iters
+    iters += 1
+
+    load_offset = tl.arange(0, bs_upper)
+    seq_lens = tl.load(paged_kernel_lens + load_offset, mask=load_offset < bid)
+    seq_len = tl.load(paged_kernel_lens + bid)
+    cum_seq_len = tl.sum(seq_lens)
+
+    kv_offset = cum_seq_len * topk + bid * iters * topk + topk_id * (seq_len + iters)
+    kv_ptr = kv_indices + kv_offset
+    token_pool_ptr = req_to_token + tl.load(req_pool_indices + bid) * pool_len
+
+    kv_offset = tl.arange(0, BLOCK_SIZE)
+    num_loop = tl.cdiv(seq_len, BLOCK_SIZE)
+    for _ in range(num_loop):
+        mask = kv_offset < seq_len
+        data = tl.load(token_pool_ptr + kv_offset, mask=mask)
+        tl.store(kv_ptr + kv_offset, data, mask=mask)
+        kv_offset += BLOCK_SIZE
+
+    extend_offset = tl.arange(0, iter_upper)
+    extend_data = tl.load(
+        token_pool_ptr + seq_len + tl.arange(0, iter_upper) * topk + topk_id,
+        mask=extend_offset < iters,
+    )
+    tl.store(kv_ptr + seq_len + extend_offset, extend_data, mask=extend_offset < iters)
+
+    # Update kv_indptr
+    bs_offset = tl.arange(0, num_tokens_upper)
+
+    zid = bid * topk + topk_id
+    if zid == 0:
+        zid = num_seqs * topk
+    positions = tl.load(positions + bs_offset, mask=bs_offset < zid)
+    base = tl.sum(positions)
+    tl.store(kv_indptr + zid, base + zid * iters)
+
+
+@torch.compile
+def select_top_k_tokens(
+    i: int,
+    topk_p: torch.Tensor,
+    topk_index: torch.Tensor,
+    hidden_states: torch.Tensor,
+    scores: torch.Tensor,
+    topk: int,
+):
+    if i == 0:
+        # The first step after extend
+        input_ids = topk_index.flatten()
+        hidden_states = hidden_states.repeat_interleave(topk, dim=0)
+        scores = topk_p  # shape: (b, topk)
+
+        tree_info = (
+            topk_p.unsqueeze(1),  # shape: (b, 1, topk)
+            topk_index,  # shape: (b, topk)
+            torch.arange(-1, topk, dtype=torch.long, device="cuda")
+            .unsqueeze(0)
+            .repeat(topk_p.shape[0], 1),  # shape: (b, topk + 1)
+        )
+
+    else:
+        # The later decode steps
+        expand_scores = torch.mul(
+            scores.unsqueeze(2), topk_p.reshape(-1, topk, topk)
+        )  # (b, topk, 1) x (b, topk ,topk) -> (b, topk, topk)
+
+        topk_cs_p, topk_cs_index = fast_topk(
+            expand_scores.flatten(start_dim=1), topk, dim=-1
+        )  # (b, topk)
+        scores = topk_cs_p  # shape: (b, topk)
+
+        topk_index = topk_index.reshape(-1, topk**2)
+        input_ids = torch.gather(topk_index, index=topk_cs_index, dim=1).flatten()
+
+        selected_input_index = topk_cs_index.flatten() // topk + torch.arange(
+            0, hidden_states.shape[0], step=topk, device="cuda"
+        ).repeat_interleave(topk)
+        hidden_states = hidden_states[selected_input_index, :]
+
+        tree_info = (
+            expand_scores,  # shape: (b, topk, topk)
+            topk_index,  # shape: (b, topk * topk)
+            topk_cs_index + (topk**2 * (i - 1) + topk),  # shape: (b, topk)
+        )
+
+    return input_ids, hidden_states, scores, tree_info
+
+
+def fast_topk(values, topk, dim):
+    if topk == 1:
+        # Use max along the specified dimension to get both value and index
+        max_value, max_index = torch.max(values, dim=dim)
+        return max_value.unsqueeze(1), max_index.unsqueeze(1)
+    else:
+        # Use topk for efficiency with larger k values
+        return torch.topk(values, topk, dim=dim)

python/sglang/srt/speculative/eagle_worker.py

+import logging
+import time
 from typing import List, Optional, Union
 
 import torch
@@ -12,8 +14,18 @@ from sglang.srt.model_executor.forward_batch_info import (
 )
 from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.speculative.eagle_utils import EAGLEDraftInput
-from sglang.srt.utils import rank0_print
+from sglang.srt.speculative.eagle_draft_cuda_graph_runner import (
+    EAGLEDraftCudaGraphRunner,
+)
+from sglang.srt.speculative.eagle_utils import (
+    EagleDraftInput,
+    EagleVerifyInput,
+    assign_draft_cache_locs,
+    fast_topk,
+    select_top_k_tokens,
+)
+
+logger = logging.getLogger(__name__)
 
 
 class EAGLEWorker(TpModelWorker):
@@ -40,41 +52,47 @@ class EAGLEWorker(TpModelWorker):
             is_draft_worker=True,
         )
         self.target_worker = target_worker
-        self.server_args = server_args
         self.finish_extend_len = []
 
+        # Parse arguments
+        self.topk = server_args.speculative_eagle_topk
+        self.speculative_num_steps = server_args.speculative_num_steps
+        self.server_args = server_args
+
         # Share the embedding and lm_head
         embed, head = self.target_worker.model_runner.model.get_embed_and_head()
         self.model_runner.model.set_embed_and_head(embed, head)
         self.model_runner.server_args.disable_cuda_graph = backup_disable_cuda_graph
-        self.model_runner.init_cuda_graphs()
 
-    def forward_draft_decode(self, batch: ScheduleBatch):
-        batch.spec_info.prepare_for_decode(batch)
-        batch.spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
-        model_worker_batch = batch.get_model_worker_batch()
-        forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
-        logits_output = self.model_runner.forward(forward_batch)
-        self.capture_for_decode(logits_output, forward_batch)
+        # Create multi-step attn backends and cuda graph runners
+        from sglang.srt.layers.attention.flashinfer_backend import (
+            FlashInferMultiStepDraftBackend,
+        )
 
-    def forward_draft_extend(self, batch: ScheduleBatch):
-        self._set_mem_pool(batch, self.model_runner)
-        batch.spec_info.prepare_for_extend(batch)
-        batch.spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
-        model_worker_batch = batch.get_model_worker_batch()
-        forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
-        logits_output = self.model_runner.forward(forward_batch)
-        self.capture_for_decode(logits_output, forward_batch)
-        self._set_mem_pool(batch, self.target_worker.model_runner)
+        self.draft_attn_backend = FlashInferMultiStepDraftBackend(
+            self.model_runner,
+            self.topk,
+            self.speculative_num_steps,
+        )
+        self.model_runner.draft_attn_backend = self.draft_attn_backend
+        self.init_cuda_graphs()
+
+    def init_cuda_graphs(self):
+        """Capture cuda graphs."""
+        self.cuda_graph_runner = None
+
+        if self.server_args.disable_cuda_graph:
+            return
+
+        tic = time.time()
+        logger.info("Capture cuda graph begin. This can take up to several minutes.")
+        self.cuda_graph_runner = EAGLEDraftCudaGraphRunner(self)
+        logger.info(f"Capture cuda graph end. Time elapsed: {time.time() - tic:.2f} s")
 
     def forward_batch_speculative_generation(self, batch: ScheduleBatch):
         if batch.forward_mode.is_decode():
             # Draft
-            self._set_mem_pool(batch, self.model_runner)
-            for i in range(self.server_args.speculative_num_steps):
-                self.forward_draft_decode(batch)
-            batch.spec_info.clear_draft_cache(batch)
-            self._set_mem_pool(batch, self.target_worker.model_runner)
+            spec_info: EagleVerifyInput = self.draft(batch)
 
             # Verify
             (
@@ -84,8 +102,7 @@ class EAGLEWorker(TpModelWorker):
                 self.finish_extend_len,
                 accept_length_cpu,
                 model_worker_batch,
-            ) = self.verify(batch)
-            next_draft_input.load_server_args(self.server_args)
+            ) = self.verify(batch, spec_info)
             batch.spec_info = next_draft_input
             # if it is None, means all requsets are finished
             if batch.spec_info.verified_id is not None:
@@ -107,29 +124,145 @@ class EAGLEWorker(TpModelWorker):
             )
 
             # Forward with the draft model.
-            spec_info = EAGLEDraftInput()
-            spec_info.load_server_args(self.server_args)
-            spec_info.hidden_states = logits_output.hidden_states
-            spec_info.verified_id = next_token_ids
-            batch.spec_info = spec_info
+            batch.spec_info = EagleDraftInput(
+                hidden_states=logits_output.hidden_states,
+                verified_id=next_token_ids,
+            )
             self.forward_draft_extend(batch)
             return logits_output, next_token_ids, model_worker_batch, 0
 
-    def verify(self, batch: ScheduleBatch):
-        verify_input = batch.spec_info.prepare_for_verify(batch)
-        verify_input.prepare_for_verify(batch)
+    def draft(self, batch: ScheduleBatch):
+        self._set_mem_pool(batch, self.model_runner)
+
+        # Parse args
+        num_seqs = batch.batch_size()
+        spec_info = batch.spec_info
+
+        # Allocate cache locations
+        out_cache_loc = batch.alloc_token_slots(
+            num_seqs * self.topk * self.speculative_num_steps
+        )
+        assign_draft_cache_locs[(num_seqs,)](
+            batch.req_pool_indices,
+            batch.req_to_token_pool.req_to_token,
+            batch.seq_lens,
+            out_cache_loc,
+            batch.req_to_token_pool.req_to_token.shape[1],
+            self.topk,
+            self.speculative_num_steps,
+        )
+
+        batch.out_cache_loc = out_cache_loc
+        batch.seq_lens_sum = torch.sum(batch.seq_lens).item()
+        spec_info.positions = batch.seq_lens.repeat_interleave(self.topk, dim=0)
+
+        # Get forward batch
+        spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
+        model_worker_batch = batch.get_model_worker_batch()
+        forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
+        can_cuda_graph = self.cuda_graph_runner and self.cuda_graph_runner.can_run(
+            forward_batch
+        )
+
+        if can_cuda_graph:
+            score_list, token_list, parents_list = self.cuda_graph_runner.replay(
+                forward_batch
+            )
+        else:
+            # Initialize attention backend
+            self.draft_attn_backend.init_forward_metadata(forward_batch)
+
+            # Run forward steps
+            score_list, token_list, parents_list = self.draft_forward(forward_batch)
+
+        ret = EagleVerifyInput.create(
+            spec_info.verified_id,
+            score_list,
+            token_list,
+            parents_list,
+            batch.seq_lens,
+            batch.seq_lens_sum,
+            self.topk,
+            self.speculative_num_steps,
+            self.server_args.speculative_num_draft_tokens,
+        )
+
+        # Free cache locations
+        batch.token_to_kv_pool.free(out_cache_loc)
+        self._set_mem_pool(batch, self.target_worker.model_runner)
+        return ret
+
+    def draft_forward(self, forward_batch: ForwardBatch):
+        # Parse args
+        spec_info = forward_batch.spec_info
+        out_cache_loc = forward_batch.out_cache_loc
+        topk_p, topk_index, hidden_states = (
+            spec_info.topk_p,
+            spec_info.topk_index,
+            spec_info.hidden_states,
+        )
+
+        # Return values
+        score_list: List[torch.Tensor] = []
+        token_list: List[torch.Tensor] = []
+        parents_list: List[torch.Tensor] = []
+
+        # Forward multiple steps
+        scores = None
+        for i in range(self.speculative_num_steps):
+            input_ids, hidden_states, scores, tree_info = select_top_k_tokens(
+                i, topk_p, topk_index, hidden_states, scores, self.topk
+            )
+            score_list.append(tree_info[0])
+            token_list.append(tree_info[1])
+            parents_list.append(tree_info[2])
+
+            # Set inputs
+            forward_batch.input_ids = input_ids
+            forward_batch.out_cache_loc = out_cache_loc[
+                forward_batch.batch_size
+                * self.topk
+                * i : forward_batch.batch_size
+                * self.topk
+                * (i + 1)
+            ]
+            forward_batch.positions.add_(1)
+            forward_batch.attn_backend = self.draft_attn_backend.attn_backends[i]
+            spec_info.hidden_states = hidden_states
+
+            # Run forward
+            logits_output = self.model_runner.model.forward(
+                forward_batch.input_ids, forward_batch.positions, forward_batch
+            )
+            probs = torch.softmax(logits_output.next_token_logits, dim=-1)
+            topk_p, topk_index = fast_topk(probs, self.topk, dim=-1)
+            hidden_states = logits_output.hidden_states
+
+        return score_list, token_list, parents_list
+
+    def verify(self, batch: ScheduleBatch, spec_info: EagleVerifyInput):
+        spec_info.prepare_for_verify(batch)
         batch.forward_mode = ForwardMode.TARGET_VERIFY
-        batch.spec_info = verify_input
-        batch.spec_info.capture_hidden_mode = CaptureHiddenMode.FULL
+        batch.spec_info = spec_info
         model_worker_batch = batch.get_model_worker_batch()
         logits_output, _ = self.target_worker.forward_batch_generation(
             model_worker_batch, skip_sample=True
         )
-        verify_input.hidden_states = logits_output.hidden_states
-        res = verify_input.verify(batch, logits_output)
+        spec_info.hidden_states = logits_output.hidden_states
+        res = spec_info.verify(batch, logits_output)
         batch.forward_mode = ForwardMode.DECODE
         return res + (model_worker_batch,)
 
+    def forward_draft_extend(self, batch: ScheduleBatch):
+        self._set_mem_pool(batch, self.model_runner)
+        batch.spec_info.prepare_for_extend(batch)
+        batch.spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
+        model_worker_batch = batch.get_model_worker_batch()
+        forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
+        logits_output = self.model_runner.forward(forward_batch)
+        self.capture_for_decode(logits_output, forward_batch)
+        self._set_mem_pool(batch, self.target_worker.model_runner)
+
     def _set_mem_pool(self, batch: ScheduleBatch, runner: ModelRunner):
         batch.token_to_kv_pool = runner.token_to_kv_pool
         batch.req_to_token_pool = runner.req_to_token_pool
@@ -139,7 +272,7 @@ class EAGLEWorker(TpModelWorker):
 
         self._set_mem_pool(batch, self.model_runner)
         batch.forward_mode = ForwardMode.DRAFT_EXTEND
-        batch.spec_info.prepare_extend_after_decode(batch)
+        batch.spec_info.prepare_extend_after_decode(batch, self.speculative_num_steps)
         batch.spec_info.capture_hidden_mode = CaptureHiddenMode.LAST
         model_worker_batch = batch.get_model_worker_batch()
         forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
@@ -155,13 +288,10 @@ class EAGLEWorker(TpModelWorker):
     def capture_for_decode(
         self, logits_output: LogitsProcessorOutput, forward_batch: ForwardBatch
     ):
-        sample_output = torch.softmax(
-            logits_output.next_token_logits, dim=-1
-        )  # TODO(kavioyu): Support more sampling methods
+        probs = torch.softmax(logits_output.next_token_logits, dim=-1)
         spec_info = forward_batch.spec_info
-        spec_info.sample_output = sample_output
+        spec_info.topk_p, spec_info.topk_index = fast_topk(probs, self.topk, dim=-1)
         spec_info.hidden_states = logits_output.hidden_states
-        spec_info.prev_mode = forward_batch.forward_mode
 
     # Don't support prefix share now.
     def finish_request(self, reqs: Union[Req, List[Req]]):