sglang 0.4.3.post3__py3-none-any.whl → 0.4.4__py3-none-any.whl

sglang/srt/model_executor/forward_batch_info.py

@@ -38,12 +38,12 @@ import triton
 import triton.language as tl
 
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
-from sglang.srt.utils import get_compiler_backend
+from sglang.srt.utils import get_compiler_backend, next_power_of_2
 
 if TYPE_CHECKING:
     from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
     from sglang.srt.managers.schedule_batch import ImageInputs, ModelWorkerBatch
-    from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
+    from sglang.srt.mem_cache.memory_pool import KVCache, ReqToTokenPool
     from sglang.srt.model_executor.model_runner import ModelRunner
     from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
     from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
@@ -51,9 +51,8 @@ if TYPE_CHECKING:
 
 
 class ForwardMode(IntEnum):
-    # Prefill a new sequence. This is deprecated now. "EXTEND" covers this case.
-    PREFILL = auto()
     # Extend a sequence. The KV cache of the beginning part of the sequence is already computed (e.g., system prompt).
+    # It is also called "prefill" in common terminology.
     EXTEND = auto()
     # Decode one token.
     DECODE = auto()
@@ -153,6 +152,12 @@ class ForwardBatch:
     top_logprobs_nums: Optional[List[int]] = None
     token_ids_logprobs: Optional[List[List[int]]] = None
 
+    # For logits and logprobs post processing
+    temp_scaled_logprobs: bool = False
+    temperature: torch.Tensor = None
+    top_p_normalized_logprobs: bool = False
+    top_p: torch.Tensor = None
+
     # Position information
     positions: torch.Tensor = None
 
@@ -189,7 +194,7 @@ class ForwardBatch:
 
     # Attention backend
     req_to_token_pool: ReqToTokenPool = None
-    token_to_kv_pool: BaseTokenToKVPool = None
+    token_to_kv_pool: KVCache = None
     attn_backend: AttentionBackend = None
 
     # For DP attention
@@ -229,7 +234,6 @@ class ForwardBatch:
             extend_input_logprob_token_ids_gpu = (
                 batch.extend_input_logprob_token_ids.to(device, non_blocking=True)
             )
-
         ret = cls(
             forward_mode=batch.forward_mode,
             batch_size=len(batch.seq_lens),
@@ -417,8 +421,8 @@ def compute_position_kernel(
     prefix_len = tl.load(extend_prefix_lens + pid) if has_prefix else 0
     seq_len = tl.load(extend_seq_lens + pid)
 
-    # TODO: optimize this?
-    cumsum_start = 0
+    # NOTE: This can be slow for large bs
+    cumsum_start = tl.cast(0, tl.int64)
     for i in range(pid):
         cumsum_start += tl.load(extend_seq_lens + i)
 

sglang/srt/model_executor/model_runner.py

@@ -35,17 +35,13 @@ from sglang.srt.distributed import (
     set_custom_all_reduce,
 )
 from sglang.srt.distributed.parallel_state import monkey_patch_vllm_parallel_state
-from sglang.srt.layers.attention.double_sparsity_backend import DoubleSparseAttnBackend
-from sglang.srt.layers.attention.flashinfer_backend import FlashInferAttnBackend
-from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
-from sglang.srt.layers.attention.torch_native_backend import TorchNativeAttnBackend
-from sglang.srt.layers.attention.triton_backend import TritonAttnBackend
 from sglang.srt.layers.dp_attention import (
     get_attention_tp_group,
     get_attention_tp_size,
     initialize_dp_attention,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.quantization import monkey_patch_isinstance_for_vllm_base_layer
 from sglang.srt.layers.sampler import Sampler
 from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
 from sglang.srt.lora.lora_manager import LoRAManager
@@ -57,9 +53,16 @@ from sglang.srt.mem_cache.memory_pool import (
     ReqToTokenPool,
     TokenToKVPoolAllocator,
 )
+from sglang.srt.mem_cache.paged_allocator import PagedTokenToKVPoolAllocator
 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.model_loader.loader import (
+    DefaultModelLoader,
+    device_loading_context,
+    get_model_loader,
+)
+from sglang.srt.model_loader.utils import set_default_torch_dtype
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.server_args import ServerArgs
@@ -77,11 +80,9 @@ from sglang.srt.utils import (
     set_cpu_offload_max_bytes,
     set_cuda_arch,
 )
-from sglang.utils import get_exception_traceback
 
 logger = logging.getLogger(__name__)
 
-
 SGLANG_CI_SMALL_KV_SIZE = os.getenv("SGLANG_CI_SMALL_KV_SIZE", None)
 UNBALANCED_MODEL_LOADING_TIMEOUT_S = 300
 
@@ -118,70 +119,22 @@ class ModelRunner:
         self.spec_algorithm = SpeculativeAlgorithm.from_string(
             server_args.speculative_algorithm
         )
+        self.page_size = server_args.page_size
         self.req_to_token_pool = req_to_token_pool
         self.token_to_kv_pool_allocator = token_to_kv_pool_allocator
 
         # Model-specific adjustment
-        if (
-            self.model_config.attention_arch == AttentionArch.MLA
-            and not self.server_args.disable_mla
-        ):
-            # TODO: add MLA optimization on CPU
-            if self.server_args.device != "cpu":
-                if server_args.enable_flashinfer_mla:
-                    logger.info(
-                        "MLA optimization is turned on. Use flashinfer mla backend."
-                    )
-                    self.server_args.attention_backend = "flashinfer_mla"
-                else:
-                    logger.info("MLA optimization is turned on. Use triton backend.")
-                    self.server_args.attention_backend = "triton"
-
-        if self.server_args.enable_double_sparsity:
-            logger.info(
-                "Double sparsity optimization is turned on. Use triton backend without CUDA graph."
-            )
-            self.server_args.attention_backend = "triton"
-            self.server_args.disable_cuda_graph = True
-            if self.server_args.ds_heavy_channel_type is None:
-                raise ValueError(
-                    "Please specify the heavy channel type for double sparsity optimization."
-                )
-            self.init_double_sparsity_channel_config(
-                self.server_args.ds_heavy_channel_type
-            )
+        self.model_specific_adjustment()
 
-        if self.is_multimodal:
-            self.mem_fraction_static *= 0.95
-            logger.info(
-                f"Automatically reduce --mem-fraction-static to {self.mem_fraction_static:.3f} "
-                f"because this is a multimodal model."
-            )
-
-            if self.model_config.hf_config.architectures == [
-                "MllamaForConditionalGeneration"
-            ]:
-                logger.info("Automatically turn off --chunked-prefill-size for mllama.")
-                server_args.chunked_prefill_size = -1
-
-            if self.model_config.hf_config.architectures == [
-                "Qwen2VLForConditionalGeneration"
-            ]:
-                # TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
-                logger.info(
-                    "Automatically turn off --chunked-prefill-size and disable radix cache for qwen2-vl."
-                )
-                server_args.chunked_prefill_size = -1
-                server_args.disable_radix_cache = True
-
-        # Global vars
         if server_args.show_time_cost:
             enable_show_time_cost()
+
         if server_args.disable_outlines_disk_cache:
             from outlines.caching import disable_cache
 
             disable_cache()
 
+        # Global vars
         global_server_args_dict.update(
             {
                 "attention_backend": server_args.attention_backend,
@@ -203,11 +156,17 @@ class ModelRunner:
             }
         )
 
+        # CPU offload
         set_cpu_offload_max_bytes(int(server_args.cpu_offload_gb * 1024**3))
 
         # Get memory before model loading
         min_per_gpu_memory = self.init_torch_distributed()
 
+        # If it is a draft model tp_group can be different.
+        self.initialize(min_per_gpu_memory)
+
+    def initialize(self, min_per_gpu_memory: float):
+        server_args = self.server_args
         self.memory_saver_adapter = TorchMemorySaverAdapter.create(
             enable=self.server_args.enable_memory_saver
         )
@@ -216,18 +175,6 @@ class ModelRunner:
         self.sampler = Sampler()
         self.load_model()
 
-        # Handle the case where some of models don't finish loading.
-        try:
-            dist.monitored_barrier(
-                group=get_tp_group().cpu_group,
-                timeout=datetime.timedelta(seconds=UNBALANCED_MODEL_LOADING_TIMEOUT_S),
-                wait_all_ranks=True,
-            )
-        except RuntimeError:
-            raise ValueError(
-                f"TP rank {self.tp_rank} could finish the model loading, but there are other ranks that didn't finish loading. It is likely due to unexpected failures (e.g., OOM) or a slow node."
-            ) from None
-
         # Apply torchao quantization
         torchao_applied = getattr(self.model, "torchao_applied", False)
         # In layered loading, torchao may have been applied
@@ -244,9 +191,11 @@ class ModelRunner:
         else:
             self.torch_tp_applied = False
 
-        # Init memory pool and attention backends
+        # Init lora
         if server_args.lora_paths is not None:
             self.init_lora_manager()
+
+        # Init memory pool and attention backends
         self.init_memory_pool(
             min_per_gpu_memory,
             server_args.max_running_requests,
@@ -260,10 +209,63 @@ class ModelRunner:
             self.cuda_graph_runner = None
             self.init_attention_backend()
 
+    def model_specific_adjustment(self):
+        server_args = self.server_args
+
+        if (
+            self.model_config.attention_arch == AttentionArch.MLA
+            and not server_args.disable_mla
+        ):
+            # TODO: add MLA optimization on CPU
+            if server_args.device != "cpu":
+                if server_args.enable_flashinfer_mla:
+                    logger.info(
+                        "MLA optimization is turned on. Use flashinfer mla backend."
+                    )
+                    server_args.attention_backend = "flashinfer_mla"
+                else:
+                    logger.info("MLA optimization is turned on. Use triton backend.")
+                    server_args.attention_backend = "triton"
+
+        if server_args.enable_double_sparsity:
+            logger.info(
+                "Double sparsity optimization is turned on. Use triton backend without CUDA graph."
+            )
+            server_args.attention_backend = "triton"
+            server_args.disable_cuda_graph = True
+            if server_args.ds_heavy_channel_type is None:
+                raise ValueError(
+                    "Please specify the heavy channel type for double sparsity optimization."
+                )
+            self.init_double_sparsity_channel_config(server_args.ds_heavy_channel_type)
+
+        if self.is_multimodal:
+            self.mem_fraction_static *= 0.95
+            logger.info(
+                f"Automatically reduce --mem-fraction-static to {self.mem_fraction_static:.3f} "
+                f"because this is a multimodal model."
+            )
+
+            if self.model_config.hf_config.architectures == [
+                "MllamaForConditionalGeneration"
+            ]:
+                logger.info("Automatically turn off --chunked-prefill-size for mllama.")
+                server_args.chunked_prefill_size = -1
+
+            if self.model_config.hf_config.architectures == [
+                "Qwen2VLForConditionalGeneration"
+            ]:
+                # TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
+                logger.info(
+                    "Automatically turn off --chunked-prefill-size and disable radix cache for qwen2-vl."
+                )
+                server_args.chunked_prefill_size = -1
+                server_args.disable_radix_cache = True
+
     def init_torch_distributed(self):
         logger.info("Init torch distributed begin.")
-        torch.get_device_module(self.device).set_device(self.gpu_id)
 
+        torch.get_device_module(self.device).set_device(self.gpu_id)
         if self.device == "cuda":
             backend = "nccl"
         elif self.device == "xpu":
@@ -304,15 +306,16 @@ class ModelRunner:
         min_per_gpu_memory = get_available_gpu_memory(
             self.device, self.gpu_id, distributed=self.tp_size > 1
         )
-        local_gpu_memory = get_available_gpu_memory(self.device, self.gpu_id)
         self.tp_group = get_tp_group()
         self.attention_tp_group = get_attention_tp_group()
 
         # Check memory for tensor parallelism
+        local_gpu_memory = get_available_gpu_memory(self.device, self.gpu_id)
         if self.tp_size > 1:
             if min_per_gpu_memory < local_gpu_memory * 0.9:
                 raise ValueError(
-                    "The memory capacity is unbalanced. Some GPUs may be occupied by other processes."
+                    "The memory capacity is unbalanced. Some GPUs may be occupied by other processes. "
+                    f"{min_per_gpu_memory=}, {local_gpu_memory=}, {local_gpu_memory * 0.9=}"
                 )
 
         logger.info(
@@ -352,6 +355,8 @@ class ModelRunner:
         # Load the model
         # Remove monkey_patch when linear.py quant remove dependencies with vllm
         monkey_patch_vllm_parallel_state()
+        monkey_patch_isinstance_for_vllm_base_layer()
+
         with self.memory_saver_adapter.region():
             self.model = get_model(
                 model_config=self.model_config,
@@ -359,6 +364,7 @@ class ModelRunner:
                 device_config=DeviceConfig(self.device),
             )
         monkey_patch_vllm_parallel_state(reverse=True)
+        monkey_patch_isinstance_for_vllm_base_layer(reverse=True)
 
         if self.server_args.kv_cache_dtype == "fp8_e4m3":
             if self.server_args.quantization_param_path is not None:
@@ -400,17 +406,22 @@ class ModelRunner:
             f"mem usage={(before_avail_memory - after_avail_memory):.2f} GB."
         )
 
+        # Handle the case where some ranks do not finish loading.
+        try:
+            dist.monitored_barrier(
+                group=get_tp_group().cpu_group,
+                timeout=datetime.timedelta(seconds=UNBALANCED_MODEL_LOADING_TIMEOUT_S),
+                wait_all_ranks=True,
+            )
+        except RuntimeError:
+            raise ValueError(
+                f"TP rank {self.tp_rank} could finish the model loading, but there are other ranks that didn't finish loading. It is likely due to unexpected failures (e.g., OOM) or a slow node."
+            ) from None
+
     def update_weights_from_disk(
         self, model_path: str, load_format: str
     ) -> tuple[bool, str]:
         """Update engine weights in-place from the disk."""
-        from sglang.srt.model_loader.loader import (
-            DefaultModelLoader,
-            device_loading_context,
-            get_model_loader,
-        )
-        from sglang.srt.model_loader.utils import set_default_torch_dtype
-
         logger.info(
             f"Update engine weights online from disk begin. "
             f"avail mem={get_available_gpu_memory(self.device, self.gpu_id):.2f} GB"
@@ -420,7 +431,7 @@ class ModelRunner:
         self.model_config.model_path = model_path
         load_config = LoadConfig(load_format=load_format)
 
-        # Only support vllm DefaultModelLoader for now
+        # Only support DefaultModelLoader for now
         loader = get_model_loader(load_config)
         if not isinstance(loader, DefaultModelLoader):
             message = f"Failed to get model loader: {loader}."
@@ -694,6 +705,12 @@ class ModelRunner:
                 )
             self.max_total_num_tokens = min(self.max_total_num_tokens, max_total_tokens)
 
+        self.max_total_num_tokens = (
+            self.max_total_num_tokens
+            // self.server_args.page_size
+            * self.server_args.page_size
+        )
+
         if self.max_total_num_tokens <= 0:
             raise RuntimeError(
                 "Not enough memory. Please try to increase --mem-fraction-static."
@@ -710,21 +727,13 @@ class ModelRunner:
             # Draft worker shares req_to_token_pool with the target worker.
             assert self.is_draft_worker
 
-        if self.token_to_kv_pool_allocator is None:
-            self.token_to_kv_pool_allocator = TokenToKVPoolAllocator(
-                self.max_total_num_tokens,
-                dtype=self.kv_cache_dtype,
-                device=self.device,
-            )
-        else:
-            assert self.is_draft_worker
-
         if (
             self.model_config.attention_arch == AttentionArch.MLA
             and not self.server_args.disable_mla
         ):
             self.token_to_kv_pool = MLATokenToKVPool(
                 self.max_total_num_tokens,
+                page_size=self.page_size,
                 dtype=self.kv_cache_dtype,
                 kv_lora_rank=self.model_config.kv_lora_rank,
                 qk_rope_head_dim=self.model_config.qk_rope_head_dim,
@@ -735,6 +744,7 @@ class ModelRunner:
         elif self.server_args.enable_double_sparsity:
             self.token_to_kv_pool = DoubleSparseTokenToKVPool(
                 self.max_total_num_tokens,
+                page_size=self.page_size,
                 dtype=self.kv_cache_dtype,
                 head_num=self.model_config.get_num_kv_heads(get_attention_tp_size()),
                 head_dim=self.model_config.head_dim,
@@ -746,6 +756,7 @@ class ModelRunner:
         else:
             self.token_to_kv_pool = MHATokenToKVPool(
                 self.max_total_num_tokens,
+                page_size=self.page_size,
                 dtype=self.kv_cache_dtype,
                 head_num=self.model_config.get_num_kv_heads(get_attention_tp_size()),
                 head_dim=self.model_config.head_dim,
@@ -753,6 +764,26 @@ class ModelRunner:
                 device=self.device,
                 enable_memory_saver=self.server_args.enable_memory_saver,
             )
+
+        if self.token_to_kv_pool_allocator is None:
+            if self.page_size == 1:
+                self.token_to_kv_pool_allocator = TokenToKVPoolAllocator(
+                    self.max_total_num_tokens,
+                    dtype=self.kv_cache_dtype,
+                    device=self.device,
+                    kvcache=self.token_to_kv_pool,
+                )
+            else:
+                self.token_to_kv_pool_allocator = PagedTokenToKVPoolAllocator(
+                    self.max_total_num_tokens,
+                    page_size=self.page_size,
+                    dtype=self.kv_cache_dtype,
+                    device=self.device,
+                    kvcache=self.token_to_kv_pool,
+                )
+        else:
+            assert self.is_draft_worker
+
         logger.info(
             f"Memory pool end. "
             f"avail mem={get_available_gpu_memory(self.device, self.gpu_id):.2f} GB"
@@ -770,6 +801,13 @@ class ModelRunner:
     def init_attention_backend(self):
         """Init attention kernel backend."""
         if self.server_args.attention_backend == "flashinfer":
+            from sglang.srt.layers.attention.flashinfer_backend import (
+                FlashInferAttnBackend,
+            )
+
+            # Init streams
+            if self.server_args.speculative_algorithm == "EAGLE":
+                self.plan_stream_for_flashinfer = torch.cuda.Stream()
             self.attn_backend = FlashInferAttnBackend(self)
         elif self.server_args.attention_backend == "triton":
             assert self.sliding_window_size is None, (
@@ -781,12 +819,26 @@ class ModelRunner:
                 "Please use `--attention-backend flashinfer`."
             )
             if self.server_args.enable_double_sparsity:
+                from sglang.srt.layers.attention.double_sparsity_backend import (
+                    DoubleSparseAttnBackend,
+                )
+
                 self.attn_backend = DoubleSparseAttnBackend(self)
             else:
+                from sglang.srt.layers.attention.triton_backend import TritonAttnBackend
+
                 self.attn_backend = TritonAttnBackend(self)
         elif self.server_args.attention_backend == "torch_native":
+            from sglang.srt.layers.attention.torch_native_backend import (
+                TorchNativeAttnBackend,
+            )
+
             self.attn_backend = TorchNativeAttnBackend(self)
         elif self.server_args.attention_backend == "flashinfer_mla":
+            from sglang.srt.layers.attention.flashinfer_mla_backend import (
+                FlashInferMLAAttnBackend,
+            )
+
             self.attn_backend = FlashInferMLAAttnBackend(self)
         else:
             raise ValueError(
@@ -878,18 +930,24 @@ class ModelRunner:
             forward_batch.input_ids, forward_batch.positions, forward_batch
         )
 
-    def forward(self, forward_batch: ForwardBatch) -> LogitsProcessorOutput:
+    def forward(
+        self, forward_batch: ForwardBatch, skip_attn_backend_init: bool = False
+    ) -> LogitsProcessorOutput:
         if (
             forward_batch.forward_mode.is_cuda_graph()
             and self.cuda_graph_runner
             and self.cuda_graph_runner.can_run(forward_batch)
         ):
-            return self.cuda_graph_runner.replay(forward_batch)
+            return self.cuda_graph_runner.replay(
+                forward_batch, skip_attn_backend_init=skip_attn_backend_init
+            )
 
         if forward_batch.forward_mode.is_decode():
             return self.forward_decode(forward_batch)
         elif forward_batch.forward_mode.is_extend():
-            return self.forward_extend(forward_batch)
+            return self.forward_extend(
+                forward_batch, skip_attn_backend_init=skip_attn_backend_init
+            )
         elif forward_batch.forward_mode.is_idle():
             return self.forward_idle(forward_batch)
         else:
@@ -909,45 +967,6 @@ class ModelRunner:
             sampling_info.update_regex_vocab_mask()
         sampling_info.apply_logits_bias(logits_output.next_token_logits)
 
-    def update_output_logprobs(
-        self,
-        logits_output: LogitsProcessorOutput,
-        sampling_info: SamplingBatchInfo,
-        top_logprobs_nums: List[int],
-        token_ids_logprobs: List[int],
-        next_token_ids: torch.Tensor,
-        *,
-        num_tokens_per_req: List[int],
-    ):
-        """Update the logits_output's output logprob based on next_token_ids
-
-        Args:
-            logits_output: The logits output from the model forward
-            sampling_info: Sampling info for logprob calculation
-            top_logprobs_nums: Number of logprobs per request.
-            next_token_ids: Next token ids.
-            num_tokens_per_req: The number of tokens per request.
-
-        Returns:
-            A list of next_token_ids
-        """
-        self._preprocess_logits(logits_output, sampling_info)
-        # We should repeat top_logprobs_nums to match num_tokens_per_req.
-        top_logprobs_nums_repeat_interleaved = []
-        token_ids_logprobs_repeat_interleaved = []
-        for num, num_tokens in zip(top_logprobs_nums, num_tokens_per_req):
-            top_logprobs_nums_repeat_interleaved.extend([num] * num_tokens)
-        for token_ids, num_tokens in zip(token_ids_logprobs, num_tokens_per_req):
-            token_ids_logprobs_repeat_interleaved.extend([token_ids] * num_tokens)
-        self.sampler(
-            logits_output,
-            sampling_info,
-            True,
-            top_logprobs_nums_repeat_interleaved,
-            token_ids_logprobs_repeat_interleaved,
-            batch_next_token_ids=next_token_ids,
-        )
-
     def sample(
         self,
         logits_output: LogitsProcessorOutput,

sglang/srt/model_loader/loader.py

@@ -48,6 +48,7 @@ from sglang.srt.model_loader.weight_utils import (
     safetensors_weights_iterator,
 )
 from sglang.srt.utils import (
+    get_bool_env_var,
     get_device_capability,
     is_pin_memory_available,
     set_weight_attrs,
@@ -197,7 +198,7 @@ class DefaultModelLoader(BaseModelLoader):
 
         Returns the path to the downloaded model, or None if the model is not
         downloaded from ModelScope."""
-        if os.environ.get("SGLANG_USE_MODELSCOPE", None) == "True":
+        if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
             # download model from ModelScope hub,
             # lazy import so that modelscope is not required for normal use.
             # pylint: disable=C.

sglang/srt/model_loader/weight_utils.py

@@ -455,7 +455,7 @@ def pt_weights_iterator(
         disable=not enable_tqdm,
         bar_format=_BAR_FORMAT,
     ):
-        state = torch.load(bin_file, map_location="cpu")
+        state = torch.load(bin_file, map_location="cpu", weights_only=True)
         yield from state.items()
         del state
         torch.cuda.empty_cache()