sglang 0.5.2rc2__py3-none-any.whl → 0.5.3.post1__py3-none-any.whl

sglang/srt/managers/schedule_batch.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import enum
+
 # Copyright 2023-2024 SGLang Team
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -35,6 +37,7 @@ import copy
 import dataclasses
 import logging
 import threading
+import time
 from enum import Enum, auto
 from http import HTTPStatus
 from itertools import chain
@@ -51,18 +54,18 @@ from sglang.srt.disaggregation.base import BaseKVSender
 from sglang.srt.disaggregation.decode_schedule_batch_mixin import (
     ScheduleBatchDisaggregationDecodeMixin,
 )
+from sglang.srt.disaggregation.utils import DisaggregationMode
 from sglang.srt.distributed.parallel_state import get_tensor_model_parallel_rank
-from sglang.srt.layers.moe import is_tbo_enabled
 from sglang.srt.mem_cache.allocator import (
     BaseTokenToKVPoolAllocator,
     SWATokenToKVPoolAllocator,
 )
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.chunk_cache import ChunkCache, SWAChunkCache
-from sglang.srt.mem_cache.lora_radix_cache import LoRAKey, LoRARadixCache
-from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
+from sglang.srt.mem_cache.memory_pool import HybridReqToTokenPool, ReqToTokenPool
+from sglang.srt.mem_cache.radix_cache import RadixKey
 from sglang.srt.mem_cache.swa_radix_cache import SWARadixCache
-from sglang.srt.metrics.collector import TimeStats
+from sglang.srt.metrics.collector import SchedulerMetricsCollector, TimeStats
 from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode, ForwardMode
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.sampling.sampling_params import SamplingParams
@@ -71,8 +74,7 @@ from sglang.srt.utils import flatten_nested_list, support_triton
 
 if TYPE_CHECKING:
     from sglang.srt.configs.model_config import ModelConfig
-    from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
-    from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
+    from sglang.srt.speculative.spec_info import SpecInput, SpeculativeAlgorithm
 
 INIT_INCREMENTAL_DETOKENIZATION_OFFSET = 5
 
@@ -87,6 +89,7 @@ GLOBAL_SERVER_ARGS_KEYS = [
     "disable_flashinfer_cutlass_moe_fp4_allgather",
     "disable_radix_cache",
     "enable_dp_lm_head",
+    "enable_fp32_lm_head",
     "flashinfer_mxfp4_moe_precision",
     "enable_flashinfer_allreduce_fusion",
     "moe_dense_tp_size",
@@ -94,20 +97,24 @@ GLOBAL_SERVER_ARGS_KEYS = [
     "ep_num_redundant_experts",
     "enable_nan_detection",
     "flashinfer_mla_disable_ragged",
-    "max_micro_batch_size",
+    "pp_max_micro_batch_size",
     "disable_shared_experts_fusion",
     "sampling_backend",
     "speculative_accept_threshold_single",
     "speculative_accept_threshold_acc",
+    "speculative_attention_mode",
     "torchao_config",
     "triton_attention_reduce_in_fp32",
     "num_reserved_decode_tokens",
     "weight_loader_disable_mmap",
     "enable_multimodal",
     "enable_symm_mem",
-    "quantization",
     "enable_custom_logit_processor",
     "disaggregation_mode",
+    "enable_deterministic_inference",
+    "nsa_prefill",
+    "nsa_decode",
+    "multi_item_scoring_delimiter",
 ]
 
 # Put some global args for easy access
@@ -408,6 +415,23 @@ class MultimodalInputs:
         # other args would be kept intact
 
 
+class RequestStage(str, enum.Enum):
+    # prefill
+    PREFILL_WAITING = "prefill_waiting"
+
+    # disaggregation prefill
+    PREFILL_PREPARE = "prefill_prepare"
+    PREFILL_BOOTSTRAP = "prefill_bootstrap"
+    PREFILL_FORWARD = "prefill_forward"
+    PREFILL_TRANSFER_KV_CACHE = "prefill_transfer_kv_cache"
+
+    # disaggregation decode
+    DECODE_PREPARE = "decode_prepare"
+    DECODE_BOOTSTRAP = "decode_bootstrap"
+    DECODE_WAITING = "decode_waiting"
+    DECODE_TRANSFERRED = "decode_transferred"
+
+
 class Req:
     """The input and output status of a request."""
 
@@ -432,8 +456,12 @@ class Req:
         bootstrap_host: Optional[str] = None,
         bootstrap_port: Optional[int] = None,
         bootstrap_room: Optional[int] = None,
+        disagg_mode: Optional[DisaggregationMode] = None,
         data_parallel_rank: Optional[int] = None,
         vocab_size: Optional[int] = None,
+        priority: Optional[int] = None,
+        metrics_collector: Optional[SchedulerMetricsCollector] = None,
+        extra_key: Optional[str] = None,
     ):
         # Input and output info
         self.rid = rid
@@ -466,6 +494,14 @@ class Req:
         self.sampling_params = sampling_params
         self.custom_logit_processor = custom_logit_processor
         self.return_hidden_states = return_hidden_states
+
+        # extra key for classifying the request (e.g. cache_salt)
+        if lora_id is not None:
+            extra_key = (
+                extra_key or ""
+            ) + lora_id  # lora_id is concatenated to the extra key
+
+        self.extra_key = extra_key
         self.lora_id = lora_id
 
         # Memory pool info
@@ -484,6 +520,7 @@ class Req:
         self.stream = stream
         self.eos_token_ids = eos_token_ids
         self.vocab_size = vocab_size
+        self.priority = priority
 
         # For incremental decoding
         # ----- | --------- read_ids -------|
@@ -503,7 +540,7 @@ class Req:
 
         # Prefix info
         # The indices to kv cache for the shared prefix.
-        self.prefix_indices: torch.Tensor = []
+        self.prefix_indices: torch.Tensor = torch.empty((0,), dtype=torch.int64)
         # Number of tokens to run prefill.
         self.extend_input_len = 0
         # The relative logprob_start_len in an extend batch
@@ -513,6 +550,8 @@ class Req:
         self.host_hit_length = 0
         # The node to lock until for swa radix tree lock ref
         self.swa_uuid_for_lock: Optional[int] = None
+        # The prefix length of the last prefix matching
+        self.last_matched_prefix_len: int = 0
 
         # Whether or not if it is chunked. It increments whenever
         # it is chunked, and decrement whenever chunked request is
@@ -561,7 +600,10 @@ class Req:
             # shape: (bs, k)
             self.output_top_logprobs_val = []
             self.output_top_logprobs_idx = []
-            self.output_token_ids_logprobs_val = []
+            # Can contain either lists or GPU tensors (delayed copy optimization for prefill-only scoring)
+            self.output_token_ids_logprobs_val: List[
+                Union[List[float], torch.Tensor]
+            ] = []
             self.output_token_ids_logprobs_idx = []
         else:
             self.output_token_logprobs_val = self.output_token_logprobs_idx = (
@@ -571,6 +613,8 @@ class Req:
             ) = None
         self.hidden_states: List[List[float]] = []
         self.hidden_states_tensor = None  # Note: use tensor instead of list to transfer hidden_states when PD + MTP
+        self.output_topk_p = None
+        self.output_topk_index = None
 
         # Embedding (return values)
         self.embedding = None
@@ -588,10 +632,10 @@ class Req:
         self.spec_verify_ct = 0
 
         # For metrics
-        self.time_stats: TimeStats = TimeStats()
+        self.metrics_collector = metrics_collector
+        self.time_stats: TimeStats = TimeStats(disagg_mode=disagg_mode)
         self.has_log_time_stats: bool = False
-        self.queue_time_start = None
-        self.queue_time_end = None
+        self.last_tic = time.monotonic()
 
         # For disaggregation
         self.bootstrap_host: str = bootstrap_host
@@ -619,6 +663,27 @@ class Req:
     def seqlen(self):
         return len(self.origin_input_ids) + len(self.output_ids)
 
+    @property
+    def is_prefill_only(self) -> bool:
+        """Check if this request is prefill-only (no token generation needed)."""
+        # NOTE: when spec is enabled, prefill_only optimizations are disabled
+        from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
+
+        spec_alg = global_server_args_dict["speculative_algorithm"]
+        return self.sampling_params.max_new_tokens == 0 and (
+            spec_alg is None or spec_alg == SpeculativeAlgorithm.NONE
+        )
+
+    def add_latency(self, stage: RequestStage):
+        if self.metrics_collector is None:
+            return
+
+        now = time.monotonic()
+        self.metrics_collector.observe_per_stage_req_latency(
+            stage.value, now - self.last_tic
+        )
+        self.last_tic = now
+
     def extend_image_inputs(self, image_inputs):
         if self.multimodal_inputs is None:
             self.multimodal_inputs = image_inputs
@@ -629,51 +694,27 @@ class Req:
         # Whether request reached finished condition
         return self.finished_reason is not None
 
-    def init_next_round_input(
-        self,
-        tree_cache: Optional[BasePrefixCache] = None,
-    ):
-        self.fill_ids = self.origin_input_ids + self.output_ids
-        if tree_cache is not None:
-            if isinstance(tree_cache, LoRARadixCache):
-                (
-                    self.prefix_indices,
-                    self.last_node,
-                    self.last_host_node,
-                    self.host_hit_length,
-                ) = tree_cache.match_prefix_with_lora_id(
-                    key=LoRAKey(
-                        lora_id=self.lora_id, token_ids=self.adjust_max_prefix_ids()
-                    ),
-                )
-            else:
-                (
-                    self.prefix_indices,
-                    self.last_node,
-                    self.last_host_node,
-                    self.host_hit_length,
-                ) = tree_cache.match_prefix(
-                    key=self.adjust_max_prefix_ids(),
-                )
-        self.extend_input_len = len(self.fill_ids) - len(self.prefix_indices)
-
-    def adjust_max_prefix_ids(self):
+    def init_next_round_input(self, tree_cache: Optional[BasePrefixCache] = None):
         self.fill_ids = self.origin_input_ids + self.output_ids
         input_len = len(self.fill_ids)
-
-        # FIXME: To work around some bugs in logprob computation, we need to ensure each
-        # request has at least one token. Later, we can relax this requirement and use `input_len`.
+        # NOTE: the matched length is at most 1 less than the input length to enable logprob computation
         max_prefix_len = input_len - 1
-
-        if self.sampling_params.max_new_tokens > 0:
-            # Need at least one token to compute logits
-            max_prefix_len = min(max_prefix_len, input_len - 1)
-
         if self.return_logprob:
             max_prefix_len = min(max_prefix_len, self.logprob_start_len)
-
         max_prefix_len = max(max_prefix_len, 0)
-        return self.fill_ids[:max_prefix_len]
+        token_ids = self.fill_ids[:max_prefix_len]
+
+        if tree_cache is not None:
+            (
+                self.prefix_indices,
+                self.last_node,
+                self.last_host_node,
+                self.host_hit_length,
+            ) = tree_cache.match_prefix(
+                key=RadixKey(token_ids=token_ids, extra_key=self.extra_key)
+            )
+            self.last_matched_prefix_len = len(self.prefix_indices)
+        self.extend_input_len = len(self.fill_ids) - len(self.prefix_indices)
 
     # Based on https://github.com/vllm-project/vllm/blob/7a64d24aad69e4d2548aa0bf528d9fe63428ab01/vllm/transformers_utils/detokenizer.py#L194-L313
     def init_incremental_detokenize(self):
@@ -684,9 +725,15 @@ class Req:
             self.surr_offset = max(
                 self.read_offset - INIT_INCREMENTAL_DETOKENIZATION_OFFSET, 0
             )
+            self.surr_and_decode_ids = (
+                self.origin_input_ids_unpadded[self.surr_offset :] + self.output_ids
+            )
+            self.cur_decode_ids_len = len(self.output_ids)
+        else:
+            self.surr_and_decode_ids.extend(self.output_ids[self.cur_decode_ids_len :])
+            self.cur_decode_ids_len = len(self.output_ids)
 
-        all_ids = self.origin_input_ids_unpadded + self.output_ids
-        return all_ids[self.surr_offset :], self.read_offset - self.surr_offset
+        return self.surr_and_decode_ids, self.read_offset - self.surr_offset
 
     def check_finished(self):
         if self.finished():
@@ -749,7 +796,7 @@ class Req:
                     return
 
     def reset_for_retract(self):
-        self.prefix_indices = []
+        self.prefix_indices = torch.empty((0,), dtype=torch.int64)
         self.last_node = None
         self.swa_uuid_for_lock = None
         self.extend_input_len = 0
@@ -781,10 +828,10 @@ class Req:
             return
 
         if self.bootstrap_room is not None:
-            prefix = f"Req Time Stats(rid={self.rid}, bootstrap_room={self.bootstrap_room}, input len={len(self.origin_input_ids)}, output len={len(self.output_ids)}, type={self.time_stats.get_type().value})"
+            prefix = f"Req Time Stats(rid={self.rid}, bootstrap_room={self.bootstrap_room}, input len={len(self.origin_input_ids)}, output len={len(self.output_ids)}, type={self.time_stats.disagg_mode_str()})"
         else:
-            prefix = f"Req Time Stats(rid={self.rid}, input len={len(self.origin_input_ids)}, output len={len(self.output_ids)}, type={self.time_stats.get_type().value})"
-        logger.info(f"{prefix}: {self.time_stats}")
+            prefix = f"Req Time Stats(rid={self.rid}, input len={len(self.origin_input_ids)}, output len={len(self.output_ids)}, type={self.time_stats.disagg_mode_str()})"
+        logger.info(f"{prefix}: {self.time_stats.convert_to_duration()}")
         self.has_log_time_stats = True
 
     def set_finish_with_abort(self, error_msg: str):
@@ -807,10 +854,6 @@ class Req:
         )
 
 
-# Batch id
-bid = 0
-
-
 @dataclasses.dataclass
 class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     """Store all information of a batch on the scheduler."""
@@ -831,15 +874,11 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     # This is an optimization to reduce the overhead of the prefill check.
     batch_is_full: bool = False
 
-    # Events
-    launch_done: Optional[threading.Event] = None
-
     # For chunked prefill in PP
     chunked_req: Optional[Req] = None
 
     # Sampling info
     sampling_info: SamplingBatchInfo = None
-    next_batch_sampling_info: SamplingBatchInfo = None
 
     # Batched arguments to model runner
     input_ids: torch.Tensor = None  # shape: [b], int64
@@ -847,6 +886,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     token_type_ids: torch.Tensor = None  # shape: [b], int64
     req_pool_indices: torch.Tensor = None  # shape: [b], int64
     seq_lens: torch.Tensor = None  # shape: [b], int64
+    seq_lens_cpu: torch.Tensor = None  # shape: [b], int64
     # The output locations of the KV cache
     out_cache_loc: torch.Tensor = None  # shape: [b], int64
     output_ids: torch.Tensor = None  # shape: [b], int64
@@ -902,7 +942,8 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
     # Speculative decoding
     spec_algorithm: SpeculativeAlgorithm = None
-    spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]] = None
+    # spec_info: Optional[SpecInput] = None
+    spec_info: Optional[SpecInput] = None
 
     # Whether to return hidden states
     return_hidden_states: bool = False
@@ -911,7 +952,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     is_prefill_only: bool = False
 
     # hicache pointer for synchronizing data loading from CPU to GPU
-    hicache_consumer_index: int = 0
+    hicache_consumer_index: int = -1
 
     @classmethod
     def init_new(
@@ -950,9 +991,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             device=req_to_token_pool.device,
             spec_algorithm=spec_algorithm,
             return_hidden_states=any(req.return_hidden_states for req in reqs),
-            is_prefill_only=all(
-                req.sampling_params.max_new_tokens == 0 for req in reqs
-            ),
+            is_prefill_only=all(req.is_prefill_only for req in reqs),
             chunked_req=chunked_req,
         )
 
@@ -962,8 +1001,11 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def is_empty(self):
         return len(self.reqs) == 0
 
-    def alloc_req_slots(self, num_reqs: int):
-        req_pool_indices = self.req_to_token_pool.alloc(num_reqs)
+    def alloc_req_slots(self, num_reqs: int, reqs: Optional[List[Req]] = None):
+        if isinstance(self.req_to_token_pool, HybridReqToTokenPool):
+            req_pool_indices = self.req_to_token_pool.alloc(num_reqs, reqs)
+        else:
+            req_pool_indices = self.req_to_token_pool.alloc(num_reqs)
         if req_pool_indices is None:
             raise RuntimeError(
                 "alloc_req_slots runs out of memory. "
@@ -1000,7 +1042,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def alloc_paged_token_slots_extend(
         self,
         prefix_lens: torch.Tensor,
+        prefix_lens_cpu: torch.Tensor,
         seq_lens: torch.Tensor,
+        seq_lens_cpu: torch.Tensor,
         last_loc: torch.Tensor,
         extend_num_tokens: int,
         backup_state: bool = False,
@@ -1008,7 +1052,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         # Over estimate the number of tokens: assume each request needs a new page.
         num_tokens = (
             extend_num_tokens
-            + len(seq_lens) * self.token_to_kv_pool_allocator.page_size
+            + len(seq_lens_cpu) * self.token_to_kv_pool_allocator.page_size
         )
         self._evict_tree_cache_if_needed(num_tokens)
 
@@ -1016,7 +1060,12 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             state = self.token_to_kv_pool_allocator.backup_state()
 
         out_cache_loc = self.token_to_kv_pool_allocator.alloc_extend(
-            prefix_lens, seq_lens, last_loc, extend_num_tokens
+            prefix_lens,
+            prefix_lens_cpu,
+            seq_lens,
+            seq_lens_cpu,
+            last_loc,
+            extend_num_tokens,
         )
         if out_cache_loc is None:
             error_msg = (
@@ -1035,6 +1084,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def alloc_paged_token_slots_decode(
         self,
         seq_lens: torch.Tensor,
+        seq_lens_cpu: torch.Tensor,
         last_loc: torch.Tensor,
         backup_state: bool = False,
     ):
@@ -1045,7 +1095,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         if backup_state:
             state = self.token_to_kv_pool_allocator.backup_state()
 
-        out_cache_loc = self.token_to_kv_pool_allocator.alloc_decode(seq_lens, last_loc)
+        out_cache_loc = self.token_to_kv_pool_allocator.alloc_decode(
+            seq_lens, seq_lens_cpu, last_loc
+        )
         if out_cache_loc is None:
             error_msg = (
                 f"Decode out of memory. Try to lower your batch size.\n"
@@ -1060,6 +1112,47 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         else:
             return out_cache_loc
 
+    def write_cache_indices(
+        self,
+        req_pool_indices: List[int],
+        prefix_lens: List[int],
+        seq_lens: List[int],
+        extend_lens: List[int],
+        out_cache_loc: torch.Tensor,
+        req_pool_indices_tensor: torch.Tensor,
+        prefix_lens_tensor: torch.Tensor,
+        seq_lens_tensor: torch.Tensor,
+        extend_lens_tensor: torch.Tensor,
+        prefix_tensors: list[torch.Tensor],
+    ):
+        if support_triton(global_server_args_dict.get("attention_backend")):
+            prefix_pointers = torch.tensor(
+                [t.data_ptr() for t in prefix_tensors], device=self.device
+            )
+            # TODO: some tensors can be reused for ForwardBatchInfo (e.g., extend_lens, cumsum_start)
+            write_req_to_token_pool_triton[(len(req_pool_indices),)](
+                self.req_to_token_pool.req_to_token,
+                req_pool_indices_tensor,
+                prefix_pointers,
+                prefix_lens_tensor,
+                seq_lens_tensor,
+                extend_lens_tensor,
+                out_cache_loc,
+                self.req_to_token_pool.req_to_token.shape[1],
+            )
+        else:
+            pt = 0
+            for i in range(len(req_pool_indices)):
+                self.req_to_token_pool.write(
+                    (req_pool_indices[i], slice(0, prefix_lens[i])),
+                    prefix_tensors[i],
+                )
+                self.req_to_token_pool.write(
+                    (req_pool_indices[i], slice(prefix_lens[i], seq_lens[i])),
+                    out_cache_loc[pt : pt + extend_lens[i]],
+                )
+                pt += extend_lens[i]
+
     def prepare_encoder_info_extend(self, input_ids: List[int], seq_lens: List[int]):
         self.encoder_lens_cpu = []
         self.encoder_cached = []
@@ -1114,6 +1207,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.seq_lens = torch.tensor(seq_lens, dtype=torch.int64).to(
             self.device, non_blocking=True
         )
+        self.seq_lens_cpu = torch.tensor(seq_lens, dtype=torch.int64)
 
         if not decoder_out_cache_loc:
             self.out_cache_loc = torch.zeros(0, dtype=torch.int64).to(
@@ -1136,10 +1230,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
     def prepare_for_extend(self):
         self.forward_mode = ForwardMode.EXTEND
 
-        # Allocate req slots
-        bs = len(self.reqs)
-        req_pool_indices = self.alloc_req_slots(bs)
-
         # Init tensors
         reqs = self.reqs
         input_ids = [r.fill_ids[len(r.prefix_indices) :] for r in reqs]
@@ -1153,21 +1243,20 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             r.token_type_ids for r in reqs if r.token_type_ids is not None
         ]
 
-        req_pool_indices_tensor = torch.tensor(req_pool_indices, dtype=torch.int64).to(
-            self.device, non_blocking=True
-        )
         input_ids_tensor = torch.tensor(
             list(chain.from_iterable(input_ids)), dtype=torch.int64
         ).to(self.device, non_blocking=True)
         seq_lens_tensor = torch.tensor(seq_lens, dtype=torch.int64).to(
             self.device, non_blocking=True
         )
+        seq_lens_cpu = torch.tensor(seq_lens, dtype=torch.int64)
         orig_seq_lens_tensor = torch.tensor(orig_seq_lens, dtype=torch.int32).to(
             self.device, non_blocking=True
         )
         prefix_lens_tensor = torch.tensor(
             prefix_lens, dtype=torch.int64, device=self.device
         )
+        prefix_lens_cpu_tensor = torch.tensor(prefix_lens, dtype=torch.int64)
 
         token_type_ids_tensor = None
         if len(token_type_ids) > 0:
@@ -1177,7 +1266,49 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
         extend_lens_tensor = seq_lens_tensor - prefix_lens_tensor
 
-        # Copy prefix and do some basic check
+        # Allocate req slots
+        bs = len(self.reqs)
+        req_pool_indices = self.alloc_req_slots(bs, self.reqs)
+        req_pool_indices_tensor = torch.tensor(req_pool_indices, dtype=torch.int64).to(
+            self.device, non_blocking=True
+        )
+
+        # Allocate memory
+        if self.token_to_kv_pool_allocator.page_size == 1:
+            out_cache_loc = self.alloc_token_slots(extend_num_tokens)
+        else:
+            last_loc = [
+                (
+                    r.prefix_indices[-1:]
+                    if len(r.prefix_indices) > 0
+                    else torch.tensor([-1], device=self.device)
+                )
+                for r in self.reqs
+            ]
+            out_cache_loc = self.alloc_paged_token_slots_extend(
+                prefix_lens_tensor,
+                prefix_lens_cpu_tensor,
+                seq_lens_tensor,
+                seq_lens_cpu,
+                torch.cat(last_loc),
+                extend_num_tokens,
+            )
+
+        # Write allocated tokens to req_to_token_pool
+        self.write_cache_indices(
+            req_pool_indices,
+            prefix_lens,
+            seq_lens,
+            extend_lens,
+            out_cache_loc,
+            req_pool_indices_tensor,
+            prefix_lens_tensor,
+            seq_lens_tensor,
+            extend_lens_tensor,
+            [r.prefix_indices for r in reqs],
+        )
+
+        # Set fields
         input_embeds = []
         extend_input_logprob_token_ids = []
         multimodal_inputs = []
@@ -1187,9 +1318,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             assert seq_len - pre_len == req.extend_input_len
 
             if pre_len > 0:
-                self.req_to_token_pool.write(
-                    (req.req_pool_idx, slice(0, pre_len)), req.prefix_indices
-                )
                 if isinstance(self.tree_cache, SWAChunkCache):
                     self.tree_cache.evict_swa(
                         req, pre_len, self.model_config.attention_chunk_size
@@ -1207,13 +1335,36 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             req.is_retracted = False
 
             # Compute the relative logprob_start_len in an extend batch
+            #
+            # Key variables:
+            # - logprob_start_len: Absolute position in full sequence where logprob computation begins
+            # - extend_logprob_start_len: Relative position within current extend batch where logprob computation begins
+            # - extend_input_len: Number of tokens that need to be processed in this extend batch
+            #   (= len(fill_ids) - len(prefix_indices), where fill_ids = origin_input_ids + output_ids
+            #    and prefix_indices are the cached/shared prefix tokens)
+            #
             if req.logprob_start_len >= pre_len:
-                req.extend_logprob_start_len = min(
-                    req.logprob_start_len - pre_len,
-                    req.extend_input_len,
-                    req.seqlen - 1,
-                )
+                # Optimization for prefill-only requests: When we only need logprobs at
+                # positions beyond the input sequence (to score next-token likelihood), skip all
+                # input logprob computation during prefill since no generation will occur.
+                if self.is_prefill_only and req.logprob_start_len == len(
+                    req.origin_input_ids
+                ):
+                    # Skip ALL input logprobs: set extend_logprob_start_len = extend_input_len
+                    req.extend_logprob_start_len = req.extend_input_len
+                else:
+                    # Convert absolute logprob_start_len to relative extend_logprob_start_len
+                    #
+                    # Example: origin_input_ids=[1,2,3,4,5] (5 tokens, positions 0-4), logprob_start_len=3
+                    # Regular logic: min(3-0, 5, 5-1) = min(3,5,4) = 3
+                    # This means: "compute logprobs from position 3 onwards in extend batch"
+                    req.extend_logprob_start_len = min(
+                        req.logprob_start_len - pre_len,
+                        req.extend_input_len,
+                        req.seqlen - 1,
+                    )
             else:
+                # logprob_start_len is before the current extend batch, so start from beginning
                 req.extend_logprob_start_len = 0
 
             if self.return_logprob:
@@ -1261,23 +1412,10 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         else:
             extend_input_logprob_token_ids = None
 
-        # Allocate memory
-        if self.token_to_kv_pool_allocator.page_size == 1:
-            out_cache_loc = self.alloc_token_slots(extend_num_tokens)
-        else:
-            last_loc = get_last_loc(
-                self.req_to_token_pool.req_to_token,
-                req_pool_indices_tensor,
-                prefix_lens_tensor,
-            )
-            out_cache_loc = self.alloc_paged_token_slots_extend(
-                prefix_lens_tensor, seq_lens_tensor, last_loc, extend_num_tokens
-            )
-
-        # Set fields
         self.input_ids = input_ids_tensor
         self.req_pool_indices = req_pool_indices_tensor
         self.seq_lens = seq_lens_tensor
+        self.seq_lens_cpu = seq_lens_cpu
         self.orig_seq_lens = orig_seq_lens_tensor
         self.out_cache_loc = out_cache_loc
         self.input_embeds = (
@@ -1306,28 +1444,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.extend_lens = extend_lens
         self.extend_input_logprob_token_ids = extend_input_logprob_token_ids
 
-        # Write to req_to_token_pool
-        if support_triton(global_server_args_dict.get("attention_backend")):
-            # TODO: some tensors can be reused for ForwardBatchInfo (e.g., extend_lens, cumsum_start)
-
-            write_req_to_token_pool_triton[(bs,)](
-                self.req_to_token_pool.req_to_token,
-                req_pool_indices_tensor,
-                prefix_lens_tensor,
-                seq_lens_tensor,
-                extend_lens_tensor,
-                out_cache_loc,
-                self.req_to_token_pool.req_to_token.shape[1],
-            )
-        else:
-            pt = 0
-            for i in range(bs):
-                self.req_to_token_pool.write(
-                    (req_pool_indices[i], slice(prefix_lens[i], seq_lens[i])),
-                    out_cache_loc[pt : pt + extend_lens[i]],
-                )
-                pt += extend_lens[i]
-
         if self.model_config.is_encoder_decoder:
             self.prepare_encoder_info_extend(input_ids, seq_lens)
 
@@ -1372,21 +1488,28 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         # TODO (lianmin): Revisit this. It should be seq_len - 1
         self.extend_logprob_start_lens.extend([0] * running_bs)
 
-    def new_page_count_next_decode(self):
+    def new_page_count_next_decode(self, selected_indices: Optional[List[int]] = None):
         page_size = self.token_to_kv_pool_allocator.page_size
+        requests = (
+            self.reqs
+            if selected_indices is None
+            else [self.reqs[i] for i in selected_indices]
+        )
         if page_size == 1:
-            return len(self.reqs)
+            return len(requests)
         # In the decoding phase, the length of a request's KV cache should be
         # the total length of the request minus 1
         return (
-            sum(1 for req in self.reqs if req.seqlen % page_size == 0)
+            sum(1 for req in requests if req.seqlen % page_size == 0)
             if self.enable_overlap
-            else sum(1 for req in self.reqs if (req.seqlen - 1) % page_size == 0)
+            else sum(1 for req in requests if (req.seqlen - 1) % page_size == 0)
         )
 
-    def check_decode_mem(self, buf_multiplier=1):
+    def check_decode_mem(
+        self, buf_multiplier=1, selected_indices: Optional[List[int]] = None
+    ):
         num_tokens = (
-            self.new_page_count_next_decode()
+            self.new_page_count_next_decode(selected_indices)
             * buf_multiplier
             * self.token_to_kv_pool_allocator.page_size
         )
@@ -1412,34 +1535,10 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 reverse=True,
             )
 
-        def get_required_tokens(num_reqs: int):
-            headroom_for_spec_decode = 0
-            if server_args.speculative_algorithm:
-                headroom_for_spec_decode += (
-                    num_reqs
-                    * server_args.speculative_eagle_topk
-                    * server_args.speculative_num_steps
-                    + num_reqs * server_args.speculative_num_draft_tokens
-                )
-            return (
-                num_reqs * global_config.retract_decode_steps + headroom_for_spec_decode
-            )
-
-        def _get_available_size():
-            if self.is_hybrid:
-                return min(
-                    self.token_to_kv_pool_allocator.full_available_size(),
-                    self.token_to_kv_pool_allocator.swa_available_size(),
-                )
-            else:
-                return self.token_to_kv_pool_allocator.available_size()
-
         retracted_reqs = []
-        seq_lens_cpu = self.seq_lens.cpu().numpy()
         first_iter = True
-        while (
-            _get_available_size() < get_required_tokens(len(sorted_indices))
-            or first_iter
+        while first_iter or (
+            not self.check_decode_mem(selected_indices=sorted_indices)
         ):
             if len(sorted_indices) == 1:
                 # Corner case: only one request left
@@ -1463,41 +1562,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             idx = sorted_indices.pop()
             req = self.reqs[idx]
             retracted_reqs.append(req)
-
-            if server_args.disaggregation_mode == "decode":
-                req.offload_kv_cache(
-                    self.req_to_token_pool, self.token_to_kv_pool_allocator
-                )
-
-            if isinstance(self.tree_cache, ChunkCache):
-                # ChunkCache does not have eviction
-                token_indices = self.req_to_token_pool.req_to_token[
-                    req.req_pool_idx, : seq_lens_cpu[idx]
-                ]
-                self.token_to_kv_pool_allocator.free(token_indices)
-                self.req_to_token_pool.free(req.req_pool_idx)
-            else:
-                # TODO: apply more fine-grained retraction
-                last_uncached_pos = (
-                    len(req.prefix_indices) // server_args.page_size
-                ) * server_args.page_size
-                token_indices = self.req_to_token_pool.req_to_token[
-                    req.req_pool_idx, last_uncached_pos : seq_lens_cpu[idx]
-                ]
-                self.token_to_kv_pool_allocator.free(token_indices)
-                self.req_to_token_pool.free(req.req_pool_idx)
-
-                # release the last node
-                if self.is_hybrid:
-                    self.tree_cache.dec_lock_ref(req.last_node, req.swa_uuid_for_lock)
-                else:
-                    self.tree_cache.dec_lock_ref(req.last_node)
-
-                # NOTE(lsyin): we should use the newly evictable memory instantly.
-                num_tokens = len(sorted_indices) * global_config.retract_decode_steps
-                self._evict_tree_cache_if_needed(num_tokens)
-
-            req.reset_for_retract()
+            self.release_req(idx, len(sorted_indices), server_args)
 
             if len(retracted_reqs) == 0:
                 # Corner case: only one request left
@@ -1516,7 +1581,45 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         ) / total_max_new_tokens
         new_estimate_ratio = min(1.0, new_estimate_ratio)
 
-        return retracted_reqs, new_estimate_ratio
+        return retracted_reqs, new_estimate_ratio, []
+
+    def release_req(self, idx: int, remaing_req_count: int, server_args: ServerArgs):
+        req = self.reqs[idx]
+        seq_lens_cpu = self.seq_lens_cpu.numpy()
+
+        if server_args.disaggregation_mode == "decode":
+            req.offload_kv_cache(
+                self.req_to_token_pool, self.token_to_kv_pool_allocator
+            )
+        if isinstance(self.tree_cache, ChunkCache):
+            # ChunkCache does not have eviction
+            token_indices = self.req_to_token_pool.req_to_token[
+                req.req_pool_idx, : seq_lens_cpu[idx]
+            ]
+            self.token_to_kv_pool_allocator.free(token_indices)
+            self.req_to_token_pool.free(req.req_pool_idx)
+        else:
+            # TODO: apply more fine-grained retraction
+            last_uncached_pos = (
+                len(req.prefix_indices) // server_args.page_size
+            ) * server_args.page_size
+            token_indices = self.req_to_token_pool.req_to_token[
+                req.req_pool_idx, last_uncached_pos : seq_lens_cpu[idx]
+            ]
+            self.token_to_kv_pool_allocator.free(token_indices)
+            self.req_to_token_pool.free(req.req_pool_idx)
+
+            # release the last node
+            if self.is_hybrid:
+                self.tree_cache.dec_lock_ref(req.last_node, req.swa_uuid_for_lock)
+            else:
+                self.tree_cache.dec_lock_ref(req.last_node)
+
+            # NOTE(lsyin): we should use the newly evictable memory instantly.
+            num_tokens = remaing_req_count * global_config.retract_decode_steps
+            self._evict_tree_cache_if_needed(num_tokens)
+
+        req.reset_for_retract()
 
     def prepare_encoder_info_decode(self):
         # Reset the encoder cached status
@@ -1526,6 +1629,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.forward_mode = ForwardMode.IDLE
         self.input_ids = torch.empty(0, dtype=torch.int64, device=self.device)
         self.seq_lens = torch.empty(0, dtype=torch.int64, device=self.device)
+        self.seq_lens_cpu = torch.empty(0, dtype=torch.int64)
         self.orig_seq_lens = torch.empty(0, dtype=torch.int32, device=self.device)
         self.out_cache_loc = torch.empty(0, dtype=torch.int64, device=self.device)
         self.req_pool_indices = torch.empty(0, dtype=torch.int32, device=self.device)
@@ -1540,7 +1644,11 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.forward_mode = ForwardMode.DECODE
         bs = len(self.reqs)
 
-        if self.spec_algorithm.is_eagle():
+        if (
+            self.spec_algorithm.is_eagle()
+            or self.spec_algorithm.is_standalone()
+            or self.spec_algorithm.is_ngram()
+        ):
             # if spec decoding is used, the decode batch is prepared inside
             # `forward_batch_speculative_generation` after running draft models.
             return
@@ -1581,10 +1689,12 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         if self.enable_overlap:
             # Do not use in-place operations in the overlap mode
             self.seq_lens = self.seq_lens + 1
+            self.seq_lens_cpu = self.seq_lens_cpu + 1
             self.orig_seq_lens = self.orig_seq_lens + 1
         else:
             # A faster in-place version
             self.seq_lens.add_(1)
+            self.seq_lens_cpu.add_(1)
             self.orig_seq_lens.add_(1)
         self.seq_lens_sum += bs
 
@@ -1603,7 +1713,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 self.req_pool_indices, self.seq_lens - 2
             ]
             self.out_cache_loc = self.alloc_paged_token_slots_decode(
-                self.seq_lens, last_loc
+                self.seq_lens, self.seq_lens_cpu, last_loc
             )
 
         self.req_to_token_pool.write(
@@ -1649,6 +1759,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             self.multimodal_inputs = [self.multimodal_inputs[i] for i in keep_indices]
         self.req_pool_indices = self.req_pool_indices[keep_indices_device]
         self.seq_lens = self.seq_lens[keep_indices_device]
+        self.seq_lens_cpu = self.seq_lens_cpu[keep_indices]
         self.orig_seq_lens = self.orig_seq_lens[keep_indices_device]
         self.out_cache_loc = None
         self.seq_lens_sum = self.seq_lens.sum().item()
@@ -1666,7 +1777,14 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
         self.sampling_info.filter_batch(keep_indices, keep_indices_device)
         if self.spec_info:
-            self.spec_info.filter_batch(keep_indices_device)
+            if chunked_req_to_exclude is not None and len(chunked_req_to_exclude) > 0:
+                has_been_filtered = False
+            else:
+                has_been_filtered = True
+            self.spec_info.filter_batch(
+                new_indices=keep_indices_device,
+                has_been_filtered=has_been_filtered,
+            )
 
     def merge_batch(self, other: "ScheduleBatch"):
         # Penalizer orchestrator must be merged before Batch.reqs is merged. This is because
@@ -1682,6 +1800,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             [self.req_pool_indices, other.req_pool_indices]
         )
         self.seq_lens = torch.cat([self.seq_lens, other.seq_lens])
+        self.seq_lens_cpu = torch.cat([self.seq_lens_cpu, other.seq_lens_cpu])
         self.orig_seq_lens = torch.cat([self.orig_seq_lens, other.orig_seq_lens])
         self.out_cache_loc = None
         self.seq_lens_sum += other.seq_lens_sum
@@ -1725,15 +1844,10 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 self.sampling_info.grammars = None
 
         seq_lens_cpu = (
-            seq_lens_cpu_cache
-            if seq_lens_cpu_cache is not None
-            else self.seq_lens.cpu()
+            seq_lens_cpu_cache if seq_lens_cpu_cache is not None else self.seq_lens_cpu
         )
 
-        global bid
-        bid += 1
         return ModelWorkerBatch(
-            bid=bid,
             forward_mode=self.forward_mode,
             input_ids=self.input_ids,
             req_pool_indices=self.req_pool_indices,
@@ -1779,7 +1893,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 )
             ),
             extend_input_logprob_token_ids=self.extend_input_logprob_token_ids,
-            launch_done=self.launch_done,
+            is_prefill_only=self.is_prefill_only,
         )
 
     def copy(self):
@@ -1852,8 +1966,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
 @dataclasses.dataclass
 class ModelWorkerBatch:
-    # The batch id
-    bid: int
     # The forward mode
     forward_mode: ForwardMode
     # The input ids
@@ -1914,19 +2026,25 @@ class ModelWorkerBatch:
 
     # Speculative decoding
     spec_algorithm: SpeculativeAlgorithm = None
-    spec_info: Optional[Union[EagleVerifyInput, EagleDraftInput]] = None
+
+    spec_info: Optional[SpecInput] = None
+
     # If set, the output of the batch contains the hidden states of the run.
     capture_hidden_mode: CaptureHiddenMode = None
-    hicache_consumer_index: int = 0
+    hicache_consumer_index: int = -1
+
+    # Overlap scheduler related
+    delay_sample_launch: bool = False
 
-    # Overlap event
-    launch_done: Optional[threading.Event] = None
+    # Whether this batch is prefill-only (no token generation needed)
+    is_prefill_only: bool = False
 
 
 @triton.jit
 def write_req_to_token_pool_triton(
     req_to_token_ptr,  # [max_batch, max_context_len]
     req_pool_indices,
+    prefix_tensors,
     pre_lens,
     seq_lens,
     extend_lens,
@@ -1939,6 +2057,19 @@ def write_req_to_token_pool_triton(
     req_pool_index = tl.load(req_pool_indices + pid)
     pre_len = tl.load(pre_lens + pid)
     seq_len = tl.load(seq_lens + pid)
+    prefix_tensor = tl.load(prefix_tensors + pid).to(tl.pointer_type(tl.int64))
+
+    # write prefix
+    num_loop = tl.cdiv(pre_len, BLOCK_SIZE)
+    for i in range(num_loop):
+        offset = tl.arange(0, BLOCK_SIZE) + i * BLOCK_SIZE
+        mask = offset < pre_len
+        value = tl.load(prefix_tensor + offset, mask=mask)
+        tl.store(
+            req_to_token_ptr + req_pool_index * req_to_token_ptr_stride + offset,
+            value,
+            mask=mask,
+        )
 
     # NOTE: This can be slow for large bs
     cumsum_start = tl.cast(0, tl.int64)