sglang 0.4.6.post1__py3-none-any.whl → 0.4.6.post2__py3-none-any.whl

sglang/srt/managers/tp_worker.py

@@ -15,11 +15,12 @@
 
 import logging
 import threading
-from typing import Optional, Tuple
+from typing import Optional, Tuple, Union
 
 import torch
 
 from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.distributed import get_pp_group, get_tp_group, get_world_group
 from sglang.srt.hf_transformers_utils import get_processor, get_tokenizer
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
 from sglang.srt.managers.io_struct import (
@@ -31,7 +32,7 @@ from sglang.srt.managers.io_struct import (
 )
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch, global_server_args_dict
 from sglang.srt.mem_cache.memory_pool import ReqToTokenPool, TokenToKVPoolAllocator
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import MultiprocessingSerializer, broadcast_pyobj, set_random_seed
@@ -47,6 +48,7 @@ class TpModelWorker:
         server_args: ServerArgs,
         gpu_id: int,
         tp_rank: int,
+        pp_rank: int,
         dp_rank: Optional[int],
         nccl_port: int,
         is_draft_worker: bool = False,
@@ -54,7 +56,9 @@ class TpModelWorker:
         token_to_kv_pool_allocator: Optional[TokenToKVPoolAllocator] = None,
     ):
         # Parse args
+        self.tp_size = server_args.tp_size
         self.tp_rank = tp_rank
+        self.pp_rank = pp_rank
 
         # Init model and tokenizer
         self.model_config = ModelConfig(
@@ -71,13 +75,17 @@ class TpModelWorker:
             enable_multimodal=server_args.enable_multimodal,
             dtype=server_args.dtype,
             quantization=server_args.quantization,
+            is_draft_model=is_draft_worker,
         )
+
         self.model_runner = ModelRunner(
             model_config=self.model_config,
             mem_fraction_static=server_args.mem_fraction_static,
             gpu_id=gpu_id,
             tp_rank=tp_rank,
             tp_size=server_args.tp_size,
+            pp_rank=pp_rank,
+            pp_size=server_args.pp_size,
             nccl_port=nccl_port,
             server_args=server_args,
             is_draft_worker=is_draft_worker,
@@ -104,6 +112,10 @@ class TpModelWorker:
                 )
         self.device = self.model_runner.device
 
+        # Init nccl groups
+        self.pp_group = get_pp_group()
+        self.world_group = get_world_group()
+
         # Profile number of tokens
         self.max_total_num_tokens = self.model_runner.max_total_num_tokens
         self.max_prefill_tokens = server_args.max_prefill_tokens
@@ -129,8 +141,9 @@ class TpModelWorker:
         # Sync random seed across TP workers
         self.random_seed = broadcast_pyobj(
             [server_args.random_seed],
-            self.tp_rank,
-            self.model_runner.tp_group.cpu_group,
+            self.tp_size * self.pp_rank + tp_rank,
+            self.world_group.cpu_group,
+            src=self.world_group.ranks[0],
         )[0]
         set_random_seed(self.random_seed)
 
@@ -155,11 +168,14 @@ class TpModelWorker:
     def get_pad_input_ids_func(self):
         return getattr(self.model_runner.model, "pad_input_ids", None)
 
-    def get_tp_cpu_group(self):
-        return self.model_runner.tp_group.cpu_group
+    def get_tp_group(self):
+        return self.model_runner.tp_group
+
+    def get_attention_tp_group(self):
+        return self.model_runner.attention_tp_group
 
     def get_attention_tp_cpu_group(self):
-        return self.model_runner.attention_tp_group.cpu_group
+        return getattr(self.model_runner.attention_tp_group, "cpu_group", None)
 
     def get_memory_pool(self):
         return (
@@ -171,19 +187,38 @@ class TpModelWorker:
         self,
         model_worker_batch: ModelWorkerBatch,
         skip_sample: bool = False,
-    ) -> Tuple[LogitsProcessorOutput, Optional[torch.Tensor]]:
+    ) -> Tuple[Union[LogitsProcessorOutput, torch.Tensor], Optional[torch.Tensor]]:
         forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)
-        logits_output = self.model_runner.forward(forward_batch)
 
-        if model_worker_batch.launch_done is not None:
-            model_worker_batch.launch_done.set()
+        pp_proxy_tensors = None
+        if not self.pp_group.is_first_rank:
+            pp_proxy_tensors = PPProxyTensors(
+                self.pp_group.recv_tensor_dict(
+                    all_gather_group=self.get_attention_tp_group()
+                )
+            )
+
+        if self.pp_group.is_last_rank:
+            logits_output = self.model_runner.forward(
+                forward_batch, pp_proxy_tensors=pp_proxy_tensors
+            )
+            if model_worker_batch.launch_done is not None:
+                model_worker_batch.launch_done.set()
 
-        if skip_sample:
-            next_token_ids = None
-        else:
-            next_token_ids = self.model_runner.sample(logits_output, model_worker_batch)
+            if skip_sample:
+                next_token_ids = None
+            else:
+                next_token_ids = self.model_runner.sample(
+                    logits_output, model_worker_batch
+                )
 
-        return logits_output, next_token_ids
+            return logits_output, next_token_ids
+        else:
+            pp_proxy_tensors = self.model_runner.forward(
+                forward_batch,
+                pp_proxy_tensors=pp_proxy_tensors,
+            )
+            return pp_proxy_tensors.tensors, None
 
     def forward_batch_embedding(self, model_worker_batch: ModelWorkerBatch):
         forward_batch = ForwardBatch.init_new(model_worker_batch, self.model_runner)

sglang/srt/managers/tp_worker_overlap_thread.py

@@ -56,11 +56,14 @@ class TpModelWorkerClient:
         server_args: ServerArgs,
         gpu_id: int,
         tp_rank: int,
+        pp_rank: int,
         dp_rank: Optional[int],
         nccl_port: int,
     ):
         # Load the model
-        self.worker = TpModelWorker(server_args, gpu_id, tp_rank, dp_rank, nccl_port)
+        self.worker = TpModelWorker(
+            server_args, gpu_id, tp_rank, pp_rank, dp_rank, nccl_port
+        )
         self.max_running_requests = self.worker.max_running_requests
         self.device = self.worker.device
         self.gpu_id = gpu_id
@@ -91,8 +94,11 @@ class TpModelWorkerClient:
     def get_pad_input_ids_func(self):
         return self.worker.get_pad_input_ids_func()
 
-    def get_tp_cpu_group(self):
-        return self.worker.get_tp_cpu_group()
+    def get_tp_group(self):
+        return self.worker.get_tp_group()
+
+    def get_attention_tp_group(self):
+        return self.worker.get_attention_tp_group()
 
     def get_attention_tp_cpu_group(self):
         return self.worker.get_attention_tp_cpu_group()

sglang/srt/mem_cache/memory_pool.py

@@ -214,6 +214,8 @@ class MHATokenToKVPool(KVCache):
         layer_num: int,
         device: str,
         enable_memory_saver: bool,
+        start_layer: Optional[int] = None,
+        end_layer: Optional[int] = None,
     ):
         self.size = size
         self.page_size = page_size
@@ -232,6 +234,8 @@ class MHATokenToKVPool(KVCache):
         self.head_dim = head_dim
         self.layer_num = layer_num
         self._create_buffers()
+        self.start_layer = start_layer or 0
+        self.end_layer = end_layer or layer_num - 1
 
         self.layer_transfer_counter = None
         self.capture_mode = False
@@ -281,6 +285,8 @@ class MHATokenToKVPool(KVCache):
 
     # for disagg
     def get_contiguous_buf_infos(self):
+        # layer_num x [seq_len, head_num, head_dim]
+        # layer_num x [page_num, page_size, head_num, head_dim]
         kv_data_ptrs = [
             self.get_key_buffer(i).data_ptr() for i in range(self.layer_num)
         ] + [self.get_value_buffer(i).data_ptr() for i in range(self.layer_num)]
@@ -320,24 +326,24 @@ class MHATokenToKVPool(KVCache):
         # transfer prepared data from host to device
         flat_data = flat_data.to(device=self.device, non_blocking=False)
         k_data, v_data = flat_data[0], flat_data[1]
-        self.k_buffer[layer_id][indices] = k_data
-        self.v_buffer[layer_id][indices] = v_data
+        self.k_buffer[layer_id - self.start_layer][indices] = k_data
+        self.v_buffer[layer_id - self.start_layer][indices] = v_data
 
     def get_key_buffer(self, layer_id: int):
         if self.layer_transfer_counter is not None:
-            self.layer_transfer_counter.wait_until(layer_id)
+            self.layer_transfer_counter.wait_until(layer_id - self.start_layer)
 
         if self.store_dtype != self.dtype:
-            return self.k_buffer[layer_id].view(self.dtype)
-        return self.k_buffer[layer_id]
+            return self.k_buffer[layer_id - self.start_layer].view(self.dtype)
+        return self.k_buffer[layer_id - self.start_layer]
 
     def get_value_buffer(self, layer_id: int):
         if self.layer_transfer_counter is not None:
-            self.layer_transfer_counter.wait_until(layer_id)
+            self.layer_transfer_counter.wait_until(layer_id - self.start_layer)
 
         if self.store_dtype != self.dtype:
-            return self.v_buffer[layer_id].view(self.dtype)
-        return self.v_buffer[layer_id]
+            return self.v_buffer[layer_id - self.start_layer].view(self.dtype)
+        return self.v_buffer[layer_id - self.start_layer]
 
     def get_kv_buffer(self, layer_id: int):
         return self.get_key_buffer(layer_id), self.get_value_buffer(layer_id)
@@ -369,12 +375,12 @@ class MHATokenToKVPool(KVCache):
             current_stream = self.device_module.current_stream()
             self.alt_stream.wait_stream(current_stream)
             with self.device_module.stream(self.alt_stream):
-                self.k_buffer[layer_id][loc] = cache_k
-            self.v_buffer[layer_id][loc] = cache_v
+                self.k_buffer[layer_id - self.start_layer][loc] = cache_k
+            self.v_buffer[layer_id - self.start_layer][loc] = cache_v
             current_stream.wait_stream(self.alt_stream)
         else:
-            self.k_buffer[layer_id][loc] = cache_k
-            self.v_buffer[layer_id][loc] = cache_v
+            self.k_buffer[layer_id - self.start_layer][loc] = cache_k
+            self.v_buffer[layer_id - self.start_layer][loc] = cache_v
 
 
 @torch.compile
@@ -484,6 +490,8 @@ class MLATokenToKVPool(KVCache):
         layer_num: int,
         device: str,
         enable_memory_saver: bool,
+        start_layer: Optional[int] = None,
+        end_layer: Optional[int] = None,
     ):
         self.size = size
         self.page_size = page_size
@@ -497,6 +505,8 @@ class MLATokenToKVPool(KVCache):
         self.kv_lora_rank = kv_lora_rank
         self.qk_rope_head_dim = qk_rope_head_dim
         self.layer_num = layer_num
+        self.start_layer = start_layer or 0
+        self.end_layer = end_layer or layer_num - 1
 
         memory_saver_adapter = TorchMemorySaverAdapter.create(
             enable=enable_memory_saver
@@ -540,19 +550,21 @@ class MLATokenToKVPool(KVCache):
 
     def get_key_buffer(self, layer_id: int):
         if self.layer_transfer_counter is not None:
-            self.layer_transfer_counter.wait_until(layer_id)
+            self.layer_transfer_counter.wait_until(layer_id - self.start_layer)
 
         if self.store_dtype != self.dtype:
-            return self.kv_buffer[layer_id].view(self.dtype)
-        return self.kv_buffer[layer_id]
+            return self.kv_buffer[layer_id - self.start_layer].view(self.dtype)
+        return self.kv_buffer[layer_id - self.start_layer]
 
     def get_value_buffer(self, layer_id: int):
         if self.layer_transfer_counter is not None:
-            self.layer_transfer_counter.wait_until(layer_id)
+            self.layer_transfer_counter.wait_until(layer_id - self.start_layer)
 
         if self.store_dtype != self.dtype:
-            return self.kv_buffer[layer_id][..., : self.kv_lora_rank].view(self.dtype)
-        return self.kv_buffer[layer_id][..., : self.kv_lora_rank]
+            return self.kv_buffer[layer_id - self.start_layer][
+                ..., : self.kv_lora_rank
+            ].view(self.dtype)
+        return self.kv_buffer[layer_id - self.start_layer][..., : self.kv_lora_rank]
 
     def get_kv_buffer(self, layer_id: int):
         return self.get_key_buffer(layer_id), self.get_value_buffer(layer_id)
@@ -568,9 +580,11 @@ class MLATokenToKVPool(KVCache):
         if cache_k.dtype != self.dtype:
             cache_k = cache_k.to(self.dtype)
         if self.store_dtype != self.dtype:
-            self.kv_buffer[layer_id][loc] = cache_k.view(self.store_dtype)
+            self.kv_buffer[layer_id - self.start_layer][loc] = cache_k.view(
+                self.store_dtype
+            )
         else:
-            self.kv_buffer[layer_id][loc] = cache_k
+            self.kv_buffer[layer_id - self.start_layer][loc] = cache_k
 
     def set_mla_kv_buffer(
         self,
@@ -605,7 +619,7 @@ class MLATokenToKVPool(KVCache):
     def transfer_per_layer(self, indices, flat_data, layer_id):
         # transfer prepared data from host to device
         flat_data = flat_data.to(device=self.device, non_blocking=False)
-        self.kv_buffer[layer_id][indices] = flat_data
+        self.kv_buffer[layer_id - self.start_layer][indices] = flat_data
 
 
 class DoubleSparseTokenToKVPool(KVCache):
@@ -620,6 +634,8 @@ class DoubleSparseTokenToKVPool(KVCache):
         device: str,
         heavy_channel_num: int,
         enable_memory_saver: bool,
+        start_layer: Optional[int] = None,
+        end_layer: Optional[int] = None,
     ):
         self.size = size
         self.page_size = page_size
@@ -657,17 +673,23 @@ class DoubleSparseTokenToKVPool(KVCache):
                 for _ in range(layer_num)
             ]
 
+        self.start_layer = start_layer or 0
+        self.end_layer = end_layer or layer_num - 1
+
     def get_key_buffer(self, layer_id: int):
-        return self.k_buffer[layer_id]
+        return self.k_buffer[layer_id - self.start_layer]
 
     def get_value_buffer(self, layer_id: int):
-        return self.v_buffer[layer_id]
+        return self.v_buffer[layer_id - self.start_layer]
 
     def get_label_buffer(self, layer_id: int):
-        return self.label_buffer[layer_id]
+        return self.label_buffer[layer_id - self.start_layer]
 
     def get_kv_buffer(self, layer_id: int):
-        return self.k_buffer[layer_id], self.v_buffer[layer_id]
+        return (
+            self.k_buffer[layer_id - self.start_layer],
+            self.v_buffer[layer_id - self.start_layer],
+        )
 
     def set_kv_buffer(
         self,
@@ -679,9 +701,9 @@ class DoubleSparseTokenToKVPool(KVCache):
     ):
         # NOTE(Andy): ignore the dtype check
         layer_id = layer.layer_id
-        self.k_buffer[layer_id][loc] = cache_k
-        self.v_buffer[layer_id][loc] = cache_v
-        self.label_buffer[layer_id][loc] = cache_label
+        self.k_buffer[layer_id - self.start_layer][loc] = cache_k
+        self.v_buffer[layer_id - self.start_layer][loc] = cache_v
+        self.label_buffer[layer_id - self.start_layer][loc] = cache_label
 
     def get_flat_data(self, indices):
         pass
@@ -930,7 +952,7 @@ class MHATokenToKVPoolHost(HostKVCache):
         return self.kv_buffer[:, :, indices]
 
     def get_flat_data_by_layer(self, indices, layer_id):
-        return self.kv_buffer[:, layer_id, indices]
+        return self.kv_buffer[:, layer_id - self.start_layer, indices]
 
     def assign_flat_data(self, indices, flat_data):
         self.kv_buffer[:, :, indices] = flat_data
@@ -955,12 +977,20 @@ class MHATokenToKVPoolHost(HostKVCache):
         for i in range(len(device_indices_cpu)):
             h_index = host_indices[i * self.page_size]
             d_index = device_indices_cpu[i]
-            device_pool.k_buffer[layer_id][d_index : d_index + self.page_size].copy_(
-                self.kv_buffer[0, layer_id, h_index : h_index + self.page_size],
+            device_pool.k_buffer[layer_id - self.start_layer][
+                d_index : d_index + self.page_size
+            ].copy_(
+                self.kv_buffer[
+                    0, layer_id - self.start_layer, h_index : h_index + self.page_size
+                ],
                 non_blocking=True,
             )
-            device_pool.v_buffer[layer_id][d_index : d_index + self.page_size].copy_(
-                self.kv_buffer[1, layer_id, h_index : h_index + self.page_size],
+            device_pool.v_buffer[layer_id - self.start_layer][
+                d_index : d_index + self.page_size
+            ].copy_(
+                self.kv_buffer[
+                    1, layer_id - self.start_layer, h_index : h_index + self.page_size
+                ],
                 non_blocking=True,
             )
 
@@ -1015,7 +1045,7 @@ class MLATokenToKVPoolHost(HostKVCache):
         return self.kv_buffer[:, indices]
 
     def get_flat_data_by_layer(self, indices, layer_id):
-        return self.kv_buffer[layer_id, indices]
+        return self.kv_buffer[layer_id - self.start_layer, indices]
 
     def assign_flat_data(self, indices, flat_data):
         self.kv_buffer[:, indices] = flat_data
@@ -1036,7 +1066,11 @@ class MLATokenToKVPoolHost(HostKVCache):
         for i in range(len(device_indices_cpu)):
             h_index = host_indices[i * self.page_size]
             d_index = device_indices_cpu[i]
-            device_pool.kv_buffer[layer_id][d_index : d_index + self.page_size].copy_(
-                self.kv_buffer[layer_id, h_index : h_index + self.page_size],
+            device_pool.kv_buffer[layer_id - self.start_layer][
+                d_index : d_index + self.page_size
+            ].copy_(
+                self.kv_buffer[
+                    layer_id - self.start_layer, h_index : h_index + self.page_size
+                ],
                 non_blocking=True,
             )

sglang/srt/model_executor/cuda_graph_runner.py

@@ -16,6 +16,7 @@
 from __future__ import annotations
 
 import bisect
+import inspect
 import os
 from contextlib import contextmanager
 from typing import TYPE_CHECKING, Callable
@@ -33,12 +34,14 @@ from sglang.srt.model_executor.forward_batch_info import (
     CaptureHiddenMode,
     ForwardBatch,
     ForwardMode,
+    PPProxyTensors,
 )
 from sglang.srt.patch_torch import monkey_patch_torch_compile
 from sglang.srt.utils import (
     get_available_gpu_memory,
     get_device_memory_capacity,
     is_hip,
+    rank0_log,
 )
 
 if TYPE_CHECKING:
@@ -135,7 +138,7 @@ def get_batch_sizes_to_capture(model_runner: ModelRunner):
 
         gpu_mem = get_device_memory_capacity()
         # Batch size of each rank will not become so large when DP is on
-        if gpu_mem is not None and gpu_mem > 81920 and server_args.dp_size == 1:
+        if gpu_mem is not None and gpu_mem > 96 * 1024:
             capture_bs += list(range(160, 257, 8))
 
     if max(capture_bs) > model_runner.req_to_token_pool.size:
@@ -188,10 +191,11 @@ class CudaGraphRunner:
         self.speculative_algorithm = model_runner.server_args.speculative_algorithm
         self.tp_size = model_runner.server_args.tp_size
         self.dp_size = model_runner.server_args.dp_size
+        self.pp_size = model_runner.server_args.pp_size
 
         # Batch sizes to capture
         self.capture_bs, self.compile_bs = get_batch_sizes_to_capture(model_runner)
-
+        rank0_log(f"Capture cuda graph bs {self.capture_bs}")
         self.capture_forward_mode = ForwardMode.DECODE
         self.capture_hidden_mode = CaptureHiddenMode.NULL
         self.num_tokens_per_bs = 1
@@ -220,6 +224,9 @@ class CudaGraphRunner:
         if self.enable_torch_compile:
             set_torch_compile_config()
 
+        if self.model_runner.server_args.lora_paths is not None:
+            self.model_runner.lora_manager.init_cuda_graph_batch_info(self.max_bs)
+
         # Graph inputs
         with torch.device("cuda"):
             self.input_ids = torch.zeros((self.max_num_token,), dtype=torch.int64)
@@ -231,6 +238,19 @@ class CudaGraphRunner:
             self.positions = torch.zeros((self.max_num_token,), dtype=torch.int64)
             self.mrope_positions = torch.zeros((3, self.max_bs), dtype=torch.int64)
 
+            # pipeline parallelism
+            if self.pp_size > 1:
+                self.pp_proxy_tensors = {
+                    "hidden_states": torch.zeros(
+                        (self.max_bs, self.model_runner.model_config.hidden_size),
+                        dtype=torch.bfloat16,
+                    ),
+                    "residual": torch.zeros(
+                        (self.max_bs, self.model_runner.model_config.hidden_size),
+                        dtype=torch.bfloat16,
+                    ),
+                }
+
             # Speculative_inference
             if (
                 model_runner.spec_algorithm.is_eagle3()
@@ -381,6 +401,12 @@ class CudaGraphRunner:
             encoder_lens = None
         mrope_positions = self.mrope_positions[:, :bs]
 
+        # pipeline parallelism
+        if self.pp_size > 1:
+            pp_proxy_tensors = PPProxyTensors(
+                {k: v[:num_tokens] for k, v in self.pp_proxy_tensors.items()}
+            )
+
         if self.enable_dp_attention or self.enable_sp_layernorm:
             self.global_num_tokens_gpu.copy_(
                 torch.tensor(
@@ -403,6 +429,13 @@ class CudaGraphRunner:
             self.capture_hidden_mode = (
                 spec_info.capture_hidden_mode if spec_info else CaptureHiddenMode.NULL
             )
+        if self.model_runner.server_args.lora_paths is not None:
+            # Currently, if the lora_path in `lora_paths` is None, the lora backend will use a
+            # different logic to handle lora, so we need to set `lora_paths` to a list of non-None
+            # values if lora is enabled.
+            lora_paths = [next(iter(self.model_runner.server_args.lora_paths))] * bs
+        else:
+            lora_paths = None
 
         forward_batch = ForwardBatch(
             forward_mode=self.capture_forward_mode,
@@ -424,8 +457,12 @@ class CudaGraphRunner:
             spec_algorithm=self.model_runner.spec_algorithm,
             spec_info=spec_info,
             capture_hidden_mode=self.capture_hidden_mode,
+            lora_paths=lora_paths,
         )
 
+        if lora_paths is not None:
+            self.model_runner.lora_manager.prepare_lora_batch(forward_batch)
+
         # Attention backend
         self.model_runner.attn_backend.init_forward_metadata_capture_cuda_graph(
             bs,
@@ -442,8 +479,20 @@ class CudaGraphRunner:
             # Clean intermediate result cache for DP attention
             forward_batch.dp_local_start_pos = forward_batch.dp_local_num_tokens = None
 
-            logits_output = forward(input_ids, forward_batch.positions, forward_batch)
-            return logits_output.next_token_logits, logits_output.hidden_states
+            kwargs = {}
+            if (
+                self.pp_size > 1
+                and "pp_proxy_tensors" in inspect.signature(forward).parameters
+            ):
+                kwargs["pp_proxy_tensors"] = pp_proxy_tensors
+
+            logits_output_or_pp_proxy_tensors = forward(
+                input_ids,
+                forward_batch.positions,
+                forward_batch,
+                **kwargs,
+            )
+            return logits_output_or_pp_proxy_tensors
 
         for _ in range(2):
             torch.cuda.synchronize()
@@ -476,7 +525,11 @@ class CudaGraphRunner:
             self.capture_hidden_mode = hidden_mode_from_spec_info
             self.capture()
 
-    def replay_prepare(self, forward_batch: ForwardBatch):
+    def replay_prepare(
+        self,
+        forward_batch: ForwardBatch,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ):
         self.recapture_if_needed(forward_batch)
 
         raw_bs = forward_batch.batch_size
@@ -505,6 +558,11 @@ class CudaGraphRunner:
                 self.seq_lens_cpu.fill_(1)
             self.seq_lens_cpu[:raw_bs].copy_(forward_batch.seq_lens_cpu)
 
+        if pp_proxy_tensors:
+            for key in self.pp_proxy_tensors.keys():
+                dim = pp_proxy_tensors[key].shape[0]
+                self.pp_proxy_tensors[key][:dim].copy_(pp_proxy_tensors[key])
+
         if self.is_encoder_decoder:
             self.encoder_lens[:raw_bs].copy_(forward_batch.encoder_lens)
         if forward_batch.mrope_positions is not None:
@@ -533,10 +591,13 @@ class CudaGraphRunner:
         self.bs = bs
 
     def replay(
-        self, forward_batch: ForwardBatch, skip_attn_backend_init: bool = False
-    ) -> LogitsProcessorOutput:
+        self,
+        forward_batch: ForwardBatch,
+        skip_attn_backend_init: bool = False,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[LogitsProcessorOutput, PPProxyTensors]:
         if not skip_attn_backend_init:
-            self.replay_prepare(forward_batch)
+            self.replay_prepare(forward_batch, pp_proxy_tensors)
         else:
             # In speculative decoding, these two fields are still needed.
             self.input_ids[: self.raw_num_token].copy_(forward_batch.input_ids)
@@ -544,17 +605,19 @@ class CudaGraphRunner:
 
         # Replay
         self.graphs[self.bs].replay()
-        next_token_logits, hidden_states = self.output_buffers[self.bs]
-
-        logits_output = LogitsProcessorOutput(
-            next_token_logits=next_token_logits[: self.raw_num_token],
-            hidden_states=(
-                hidden_states[: self.raw_num_token]
-                if hidden_states is not None
-                else None
-            ),
-        )
-        return logits_output
+        output = self.output_buffers[self.bs]
+        if isinstance(output, LogitsProcessorOutput):
+            return LogitsProcessorOutput(
+                next_token_logits=output.next_token_logits[: self.raw_num_token],
+                hidden_states=(
+                    output.hidden_states[: self.raw_num_token]
+                    if output.hidden_states is not None
+                    else None
+                ),
+            )
+        else:
+            assert isinstance(output, PPProxyTensors)
+            return PPProxyTensors({k: v[: self.bs] for k, v in output.tensors.items()})
 
     def get_spec_info(self, num_tokens: int):
         spec_info = None

sglang/srt/model_executor/forward_batch_info.py

@@ -31,7 +31,7 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 from enum import IntEnum, auto
-from typing import TYPE_CHECKING, List, Optional, Union
+from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
 import torch
 import triton
@@ -585,6 +585,36 @@ class ForwardBatch:
         self.prepare_chunked_kv_indices(device)
 
 
+class PPProxyTensors:
+    # adapted from https://github.com/vllm-project/vllm/blob/d14e98d924724b284dc5eaf8070d935e214e50c0/vllm/sequence.py#L1103
+    tensors: Dict[str, torch.Tensor]
+
+    def __init__(self, tensors):
+        # manually define this function, so that
+        # Dynamo knows `IntermediateTensors()` comes from this file.
+        # Otherwise, dataclass will generate this function by evaluating
+        # a string, and we will lose the information about the source file.
+        self.tensors = tensors
+
+    def __getitem__(self, key: Union[str, slice]):
+        if isinstance(key, str):
+            return self.tensors[key]
+        elif isinstance(key, slice):
+            return self.__class__({k: v[key] for k, v in self.tensors.items()})
+
+    def __setitem__(self, key: str, value: torch.Tensor):
+        self.tensors[key] = value
+
+    def __len__(self):
+        return len(self.tensors)
+
+    def __eq__(self, other: object):
+        return isinstance(other, self.__class__) and self
+
+    def __repr__(self) -> str:
+        return f"PPProxyTensors(tensors={self.tensors})"
+
+
 def compute_position_triton(
     extend_prefix_lens: torch.Tensor, extend_seq_lens: torch.Tensor, extend_seq_lens_sum
 ):