sglang 0.4.8.post1__py3-none-any.whl → 0.4.9.post1__py3-none-any.whl

sglang/srt/server_args.py

@@ -20,7 +20,7 @@ import logging
 import os
 import random
 import tempfile
-from typing import List, Literal, Optional
+from typing import List, Literal, Optional, Union
 
 from sglang.srt.hf_transformers_utils import check_gguf_file, get_config
 from sglang.srt.reasoning_parser import ReasoningParser
@@ -46,6 +46,7 @@ class ServerArgs:
     tokenizer_path: Optional[str] = None
     tokenizer_mode: str = "auto"
     skip_tokenizer_init: bool = False
+    skip_server_warmup: bool = False
     load_format: str = "auto"
     model_loader_extra_config: str = "{}"
     trust_remote_code: bool = False
@@ -61,11 +62,13 @@ class ServerArgs:
     is_embedding: bool = False
     enable_multimodal: Optional[bool] = None
     revision: Optional[str] = None
+    hybrid_kvcache_ratio: Optional[float] = None
     impl: str = "auto"
 
     # Port for the HTTP server
     host: str = "127.0.0.1"
     port: int = 30000
+    nccl_port: Optional[int] = None
 
     # Memory and scheduling
     mem_fraction_static: Optional[float] = None
@@ -98,6 +101,7 @@ class ServerArgs:
     log_level_http: Optional[str] = None
     log_requests: bool = False
     log_requests_level: int = 0
+    crash_dump_folder: Optional[str] = None
     show_time_cost: bool = False
     enable_metrics: bool = False
     bucket_time_to_first_token: Optional[List[float]] = None
@@ -129,7 +133,7 @@ class ServerArgs:
     preferred_sampling_params: Optional[str] = None
 
     # LoRA
-    lora_paths: Optional[List[str]] = None
+    lora_paths: Optional[Union[dict[str, str], List[str]]] = None
     max_loras_per_batch: int = 8
     lora_backend: str = "triton"
 
@@ -154,6 +158,7 @@ class ServerArgs:
     enable_ep_moe: bool = False
     enable_deepep_moe: bool = False
     enable_flashinfer_moe: bool = False
+    enable_flashinfer_allreduce_fusion: bool = False
     deepep_mode: Optional[Literal["auto", "normal", "low_latency"]] = "auto"
     ep_num_redundant_experts: int = 0
     ep_dispatch_algorithm: Optional[Literal["static", "dynamic", "fake"]] = None
@@ -212,11 +217,13 @@ class ServerArgs:
     hicache_ratio: float = 2.0
     hicache_size: int = 0
     hicache_write_policy: str = "write_through_selective"
+    hicache_io_backend: str = ""
     flashinfer_mla_disable_ragged: bool = False
     disable_shared_experts_fusion: bool = False
     disable_chunked_prefix_cache: bool = False
     disable_fast_image_processor: bool = False
     enable_return_hidden_states: bool = False
+    enable_triton_kernel_moe: bool = False
     warmups: Optional[str] = None
 
     # Debug tensor dumps
@@ -315,6 +322,14 @@ class ServerArgs:
             else:
                 self.mem_fraction_static = 0.88
 
+            # Lazy init to avoid circular import
+            from sglang.srt.configs.model_config import ModelConfig
+
+            # Multimodal models need more memory for the image processor
+            model_config = ModelConfig.from_server_args(self)
+            if model_config.is_multimodal:
+                self.mem_fraction_static *= 0.90
+
         # Set chunked prefill size, which depends on the gpu memory capacity
         if self.chunked_prefill_size is None:
             if gpu_mem is not None:
@@ -376,6 +391,12 @@ class ServerArgs:
             )
             self.disable_cuda_graph = True
 
+        if self.attention_backend == "ascend":
+            logger.warning(
+                "At this moment Ascend attention backend only supports a page_size of 128, change page_size to 128."
+            )
+            self.page_size = 128
+
         # Choose grammar backend
         if self.grammar_backend is None:
             self.grammar_backend = "xgrammar"
@@ -399,10 +420,6 @@ class ServerArgs:
 
         # DeepEP MoE
         if self.enable_deepep_moe:
-            if self.deepep_mode == "auto":
-                assert (
-                    not self.enable_dp_attention
-                ), "DeepEP MoE `auto` mode is not supported with DP Attention."
             if self.deepep_mode == "normal":
                 logger.warning("Cuda graph is disabled because deepep_mode=`normal`")
                 self.disable_cuda_graph = True
@@ -485,12 +502,6 @@ class ServerArgs:
                     self.speculative_num_draft_tokens,
                 ) = auto_choose_speculative_params(self)
 
-            if self.page_size > 1 and self.speculative_eagle_topk > 1:
-                self.speculative_eagle_topk = 1
-                logger.warning(
-                    "speculative_eagle_topk is adjusted to 1 when page_size > 1"
-                )
-
             if (
                 self.speculative_eagle_topk == 1
                 and self.speculative_num_draft_tokens != self.speculative_num_steps + 1
@@ -587,6 +598,12 @@ class ServerArgs:
             default=ServerArgs.port,
             help="The port of the HTTP server.",
         )
+        parser.add_argument(
+            "--nccl-port",
+            type=int,
+            default=ServerArgs.nccl_port,
+            help="The port for NCCL distributed environment setup. Defaults to a random port.",
+        )
         parser.add_argument(
             "--tokenizer-mode",
             type=str,
@@ -601,6 +618,11 @@ class ServerArgs:
             action="store_true",
             help="If set, skip init tokenizer and pass input_ids in generate request.",
         )
+        parser.add_argument(
+            "--skip-server-warmup",
+            action="store_true",
+            help="If set, skip warmup.",
+        )
         parser.add_argument(
             "--load-format",
             type=str,
@@ -686,6 +708,7 @@ class ServerArgs:
                 "w8a8_fp8",
                 "moe_wna16",
                 "qoq",
+                "w4afp8",
             ],
             help="The quantization method.",
         )
@@ -817,6 +840,18 @@ class ServerArgs:
             default=ServerArgs.page_size,
             help="The number of tokens in a page.",
         )
+        parser.add_argument(
+            "--hybrid-kvcache-ratio",
+            nargs="?",
+            const=0.5,
+            type=float,
+            default=ServerArgs.hybrid_kvcache_ratio,
+            help=(
+                "Mix ratio in [0,1] between uniform and hybrid kv buffers "
+                "(0.0 = pure uniform: swa_size / full_size = 1)"
+                "(1.0 = pure hybrid: swa_size / full_size = local_attention_size / context_length)"
+            ),
+        )
 
         # Other runtime options
         parser.add_argument(
@@ -920,8 +955,14 @@ class ServerArgs:
             "--log-requests-level",
             type=int,
             default=0,
-            help="0: Log metadata. 1. Log metadata and partial input/output. 2. Log every input/output.",
-            choices=[0, 1, 2],
+            help="0: Log metadata (no sampling parameters). 1: Log metadata and sampling parameters. 2: Log metadata, sampling parameters and partial input/output. 3: Log every input/output.",
+            choices=[0, 1, 2, 3],
+        )
+        parser.add_argument(
+            "--crash-dump-folder",
+            type=str,
+            default=ServerArgs.crash_dump_folder,
+            help="Folder path to dump requests from the last 5 min before a crash (if any). If not specified, crash dumping is disabled.",
         )
         parser.add_argument(
             "--show-time-cost",
@@ -1092,6 +1133,7 @@ class ServerArgs:
                 "flashmla",
                 "intel_amx",
                 "torch_native",
+                "ascend",
                 "triton",
             ],
             default=ServerArgs.attention_backend,
@@ -1186,6 +1228,11 @@ class ServerArgs:
             action="store_true",
             help="Enable FlashInfer CUTLASS MoE backend for modelopt_fp4 quant on Blackwell. Supports MoE-EP with --enable-ep-moe",
         )
+        parser.add_argument(
+            "--enable-flashinfer-allreduce-fusion",
+            action="store_true",
+            help="Enable FlashInfer allreduce fusion for Add_RMSNorm.",
+        )
         parser.add_argument(
             "--enable-deepep-moe",
             action="store_true",
@@ -1485,6 +1532,13 @@ class ServerArgs:
             default=ServerArgs.hicache_write_policy,
             help="The write policy of hierarchical cache.",
         )
+        parser.add_argument(
+            "--hicache-io-backend",
+            type=str,
+            choices=["direct", "kernel"],
+            default=ServerArgs.hicache_io_backend,
+            help="The IO backend for KV cache transfer between CPU and GPU",
+        )
         parser.add_argument(
             "--flashinfer-mla-disable-ragged",
             action="store_true",
@@ -1510,6 +1564,11 @@ class ServerArgs:
             action="store_true",
             help="Enable returning hidden states with responses.",
         )
+        parser.add_argument(
+            "--enable-triton-kernel-moe",
+            action="store_true",
+            help="Use triton moe grouped gemm kernel.",
+        )
         parser.add_argument(
             "--warmups",
             type=str,
@@ -1706,14 +1765,17 @@ class PortArgs:
 
     @staticmethod
     def init_new(server_args, dp_rank: Optional[int] = None) -> "PortArgs":
-        port = server_args.port + random.randint(100, 1000)
-        while True:
-            if is_port_available(port):
-                break
-            if port < 60000:
-                port += 42
-            else:
-                port -= 43
+        if server_args.nccl_port is None:
+            port = server_args.port + random.randint(100, 1000)
+            while True:
+                if is_port_available(port):
+                    break
+                if port < 60000:
+                    port += 42
+                else:
+                    port -= 43
+        else:
+            port = server_args.nccl_port
 
         if not server_args.enable_dp_attention:
             # Normal case, use IPC within a single node

sglang/srt/speculative/build_eagle_tree.py

@@ -1,10 +1,12 @@
 # NOTE: Please run this file to make sure the test cases are correct.
 
-from typing import List
+import math
+from enum import IntEnum
+from typing import List, Optional
 
 import torch
 
-from sglang.srt.utils import is_cuda, is_hip, rank0_print
+from sglang.srt.utils import is_cuda, is_hip
 
 if is_cuda() or is_hip():
     from sgl_kernel import (
@@ -40,6 +42,12 @@ def build_tree_kernel_efficient_preprocess(
     return parent_list, top_scores_index, draft_tokens
 
 
+class TreeMaskMode(IntEnum):
+    FULL_MASK = 0
+    QLEN_ONLY = 1
+    QLEN_ONLY_BITPACKING = 2
+
+
 def build_tree_kernel_efficient(
     verified_id: torch.Tensor,
     score_list: List[torch.Tensor],
@@ -50,6 +58,9 @@ def build_tree_kernel_efficient(
     topk: int,
     spec_steps: int,
     num_verify_tokens: int,
+    tree_mask_mode: TreeMaskMode = TreeMaskMode.FULL_MASK,
+    tree_mask_buf: Optional[torch.Tensor] = None,
+    position_buf: Optional[torch.Tensor] = None,
 ):
     parent_list, top_scores_index, draft_tokens = (
         build_tree_kernel_efficient_preprocess(
@@ -66,15 +77,37 @@ def build_tree_kernel_efficient(
     device = seq_lens.device
     # e.g. for bs=1, tree_mask: num_draft_token, seq_lens_sum + num_draft_token (flattened)
     # where each row indicates the attending pattern of each draft token
+    # if use_partial_packed_tree_mask is True, tree_mask: num_draft_token (flattened, packed)
+    if tree_mask_buf is not None:
+        tree_mask = tree_mask_buf
+    elif tree_mask_mode == TreeMaskMode.QLEN_ONLY:
+        tree_mask = torch.full(
+            (num_verify_tokens * bs * num_verify_tokens,),
+            True,
+            dtype=torch.bool,
+            device=device,
+        )
+    elif tree_mask_mode == TreeMaskMode.QLEN_ONLY_BITPACKING:
+        packed_dtypes = [torch.uint8, torch.uint16, torch.uint32]
+        packed_dtype_idx = int(math.ceil(math.log2((num_verify_tokens + 7) // 8)))
+        tree_mask = torch.zeros(
+            (num_verify_tokens * bs,),
+            dtype=packed_dtypes[packed_dtype_idx],
+            device=device,
+        )
+    elif tree_mask_mode == TreeMaskMode.FULL_MASK:
+        tree_mask = torch.full(
+            (
+                seq_lens_sum * num_verify_tokens
+                + num_verify_tokens * num_verify_tokens * bs,
+            ),
+            True,
+            device=device,
+        )
+    else:
+        raise NotImplementedError(f"Invalid tree mask: {tree_mask_mode=}")
+
     # TODO: make them torch.empty and fuse them into `sgl_build_tree_kernel`
-    tree_mask = torch.full(
-        (
-            seq_lens_sum * num_verify_tokens
-            + num_verify_tokens * num_verify_tokens * bs,
-        ),
-        True,
-        device=device,
-    )
     retrive_index = torch.full(
         (bs, num_verify_tokens), -1, device=device, dtype=torch.long
     )
@@ -87,7 +120,12 @@ def build_tree_kernel_efficient(
     # position: where each token belongs to
     # e.g. if depth of each draft token is [0, 1, 1, 2] and the prompt length is 7
     # then, positions = [7, 8, 8, 9]
-    positions = torch.empty((bs * num_verify_tokens,), device=device, dtype=torch.long)
+    if position_buf is not None:
+        positions = position_buf
+    else:
+        positions = torch.empty(
+            (bs * num_verify_tokens,), device=device, dtype=torch.long
+        )
 
     sgl_build_tree_kernel_efficient(
         parent_list,
@@ -101,6 +139,7 @@ def build_tree_kernel_efficient(
         topk,
         spec_steps,
         num_verify_tokens,
+        tree_mask_mode,
     )
     return (
         tree_mask,
@@ -344,13 +383,13 @@ def test_build_tree_kernel_efficient():
         num_verify_tokens=num_draft_token,
     )
 
-    rank0_print("=========== build tree kernel efficient ==========")
-    # rank0_print(f"{tree_mask=}", flush=True)
-    rank0_print(f"{position=}", flush=True)
-    rank0_print(f"{retrive_index=}", flush=True)
-    rank0_print(f"{retrive_next_token=}", flush=True)
-    rank0_print(f"{retrive_next_sibling=}", flush=True)
-    rank0_print(f"{draft_tokens=}", flush=True)
+    print("=========== build tree kernel efficient ==========")
+    print(f"{tree_mask=}")
+    print(f"{position=}")
+    print(f"{retrive_index=}")
+    print(f"{retrive_next_token=}")
+    print(f"{retrive_next_sibling=}")
+    print(f"{draft_tokens=}")
     assert position.tolist() == [5, 6, 6, 7, 7, 8, 8, 9, 10, 11, 12, 12, 12, 12, 13, 14]
     assert retrive_index.tolist() == [
         [0, 1, 2, 3, 4, 5, 6, 7],

sglang/srt/speculative/eagle_worker.py

@@ -140,9 +140,11 @@ class EAGLEWorker(TpModelWorker):
             self.draft_model_runner.model.set_embed(embed)
 
             # grab hot token ids
-            self.hot_token_id = self.draft_model_runner.model.get_hot_token_id().to(
-                embed.device
-            )
+            if self.draft_model_runner.model.hot_token_id is not None:
+                self.hot_token_id = self.draft_model_runner.model.hot_token_id.to(
+                    embed.device
+                )
+
         else:
             if self.hot_token_id is not None:
                 head = head.clone()
@@ -842,7 +844,7 @@ class EAGLEWorker(TpModelWorker):
                 )
         batch.return_hidden_states = False
         model_worker_batch = batch.get_model_worker_batch()
-        model_worker_batch.spec_num_draft_tokens = self.speculative_num_draft_tokens
+        model_worker_batch.spec_num_draft_tokens = self.speculative_num_steps + 1
         assert model_worker_batch.capture_hidden_mode == CaptureHiddenMode.LAST
         forward_batch = ForwardBatch.init_new(
             model_worker_batch, self.draft_model_runner