sglang 0.5.2rc2__py3-none-any.whl → 0.5.3__py3-none-any.whl

sglang/srt/layers/attention/tbo_backend.py

@@ -1,10 +1,10 @@
-from typing import TYPE_CHECKING, Callable, List, Optional, Union
+from typing import TYPE_CHECKING, Callable, List, Optional
 
 import torch
 
 from sglang.srt import two_batch_overlap
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
-from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
+from sglang.srt.speculative.spec_info import SpecInput
 
 if TYPE_CHECKING:
     from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
@@ -46,7 +46,7 @@ class TboAttnBackend(AttentionBackend):
         seq_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: "ForwardMode",
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
     ):
         self.primary.init_forward_metadata_capture_cuda_graph(
             bs=bs,
@@ -77,7 +77,7 @@ class TboAttnBackend(AttentionBackend):
         seq_lens_sum: int,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: "ForwardMode",
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
     ):
         self.primary.init_forward_metadata_replay_cuda_graph(
@@ -112,7 +112,7 @@ class TboAttnBackend(AttentionBackend):
         seq_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: "ForwardMode",
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
         # capture args
         capture_num_tokens: int = None,
         # replay args
@@ -196,7 +196,7 @@ def _init_forward_metadata_cuda_graph_split(
     seq_lens: torch.Tensor,
     encoder_lens: Optional[torch.Tensor],
     forward_mode: "ForwardMode",
-    spec_info: Optional[EagleVerifyInput],
+    spec_info: Optional[SpecInput],
     # capture args
     capture_num_tokens: int = None,
     # replay args

sglang/srt/layers/attention/torch_flex_backend.py

@@ -0,0 +1,325 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+from torch.nn.attention.flex_attention import create_block_mask, flex_attention
+
+from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
+from sglang.srt.layers.radix_attention import AttentionType
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.radix_attention import RadixAttention
+    from sglang.srt.model_executor.model_runner import ModelRunner
+
+
+class TorchFlexAttnBackend(AttentionBackend):
+    def __init__(self, model_runner: ModelRunner):
+        super().__init__()
+        self.forward_metadata = None
+        self.device = model_runner.device
+        self.flex_attention = torch.compile(flex_attention, dynamic=True)
+        torch._dynamo.config.cache_size_limit = 1024
+        torch._dynamo.config.accumulated_cache_size_limit = 1024
+
+    def init_forward_metadata(self, forward_batch: ForwardBatch):
+        """Init the metadata for a forward pass."""
+        # TODO: find a more elegant way to save memory
+        # Currently maintain the same memory as torch_native_backend
+        torch.cuda.empty_cache()
+
+        # Provide two block_mask Lists per seq_idx for lower latency, later will support per layer level mask generation
+        self.extend_block_masks = []
+        self.decode_block_masks = []
+
+        if forward_batch.forward_mode.is_extend():
+            for seq_idx in range(forward_batch.seq_lens.shape[0]):
+                seq_len_kv = forward_batch.seq_lens[seq_idx]
+                seq_len_q = seq_len_kv
+                self.extend_block_masks.append(
+                    create_block_mask(
+                        self._causal_mask,
+                        None,
+                        None,
+                        seq_len_q,
+                        seq_len_kv,
+                        device=self.device,
+                        _compile=False,
+                    )
+                )
+
+        elif forward_batch.forward_mode.is_decode():
+            for seq_idx in range(forward_batch.seq_lens.shape[0]):
+                seq_len_q = 1
+                seq_len_kv = forward_batch.seq_lens[seq_idx]
+
+                self.decode_block_masks.append(
+                    create_block_mask(
+                        self._decode_mask,
+                        None,
+                        None,
+                        seq_len_q,
+                        seq_len_kv,
+                        device=self.device,
+                        _compile=False,
+                    )
+                )
+
+    def _causal_mask(self, b, h, q_idx, kv_idx):
+        return q_idx >= kv_idx
+
+    def _decode_mask(self, b, h, q_idx, kv_idx):
+        return q_idx <= kv_idx
+
+    def _run_flex_forward_extend(
+        self,
+        query: torch.Tensor,
+        output: torch.Tensor,
+        k_cache: torch.Tensor,
+        v_cache: torch.Tensor,
+        req_to_token: torch.Tensor,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        extend_prefix_lens: torch.Tensor,
+        extend_seq_lens: torch.Tensor,
+        scaling=None,
+        enable_gqa=False,
+        causal=False,
+    ):
+        """Run the extend forward by using torch flex attention op.
+
+        Args:
+            query: [num_tokens, num_heads, head_size]
+            output: [num_tokens, num_heads, head_size]
+            k_cache: [max_total_num_tokens, num_heads, head_size]
+            v_cache: [max_total_num_tokens, num_heads, head_size]
+            req_to_token: [max_num_reqs, max_context_len]
+            req_pool_indices: [num_seqs]
+            seq_lens: [num_seqs]
+            extend_prefix_lens: [num_seqs]
+            extend_seq_lens: [num_seqs]
+            scaling: float or None
+            enable_gqa: bool
+            causal: bool
+
+        Returns:
+            output: [num_tokens, num_heads, head_size]
+        """
+
+        assert seq_lens.shape[0] == extend_prefix_lens.shape[0]
+        assert seq_lens.shape[0] == extend_seq_lens.shape[0]
+
+        # [num_tokens, num_heads, head_size] -> [num_heads, num_tokens, head_size]
+        query = query.movedim(0, query.dim() - 2)
+
+        start_q, start_kv = 0, 0
+
+        for seq_idx in range(seq_lens.shape[0]):
+            # TODO: this loop process a sequence per iter, this is inefficient.
+            # Need optimize the performance later.
+            extend_seq_len_q = extend_seq_lens[seq_idx]
+            prefill_seq_len_q = extend_prefix_lens[seq_idx]
+
+            seq_len_kv = seq_lens[seq_idx]
+            end_q = start_q + extend_seq_len_q
+            end_kv = start_kv + seq_len_kv
+
+            per_req_query = query[:, start_q:end_q, :]
+            per_req_query_redundant = torch.empty(
+                (per_req_query.shape[0], seq_len_kv, per_req_query.shape[2]),
+                dtype=per_req_query.dtype,
+                device=per_req_query.device,
+            )
+
+            per_req_query_redundant[:, prefill_seq_len_q:, :] = per_req_query
+
+            # get key and value from cache. per_req_tokens contains the kv cache
+            # index for each token in the sequence.
+            req_pool_idx = req_pool_indices[seq_idx]
+            per_req_tokens = req_to_token[req_pool_idx, :seq_len_kv]
+            per_req_key = k_cache[per_req_tokens].movedim(0, query.dim() - 2)
+            per_req_value = v_cache[per_req_tokens].movedim(0, query.dim() - 2)
+
+            if not causal:
+                raise NotImplementedError("Non-causal mode is not yet implemented.")
+
+            per_req_out_redundant = (
+                self.flex_attention(
+                    per_req_query_redundant.unsqueeze(0),
+                    per_req_key.unsqueeze(0),
+                    per_req_value.unsqueeze(0),
+                    block_mask=self.extend_block_masks[seq_idx],
+                    scale=scaling,
+                    enable_gqa=enable_gqa,
+                )
+                .squeeze(0)
+                .movedim(query.dim() - 2, 0)
+            )
+            output[start_q:end_q, :, :] = per_req_out_redundant[
+                prefill_seq_len_q:, :, :
+            ]
+            start_q, start_kv = end_q, end_kv
+        return output
+
+    def _run_flex_forward_decode(
+        self,
+        query: torch.Tensor,
+        output: torch.Tensor,
+        k_cache: torch.Tensor,
+        v_cache: torch.Tensor,
+        req_to_token: torch.Tensor,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        scaling=None,
+        enable_gqa=False,
+        causal=False,
+    ):
+        """Run the decode forward by using torch flex attention op.
+
+        Args:
+            query: [num_tokens, num_heads, head_size]
+            output: [num_tokens, num_heads, head_size]
+            k_cache: [max_total_num_tokens, num_heads, head_size]
+            v_cache: [max_total_num_tokens, num_heads, head_size]
+            req_to_token: [max_num_reqs, max_context_len]
+            req_pool_indices: [num_seqs]
+            seq_lens: [num_seqs]
+            scaling: float or None
+            enable_gqa: bool
+            causal: bool
+
+        Returns:
+            output: [num_tokens, num_heads, head_size]
+        """
+
+        # [num_tokens, num_heads, head_size] -> [num_heads, num_tokens, head_size]
+        query = query.movedim(0, query.dim() - 2)
+
+        start_q, start_kv = 0, 0
+        for seq_idx in range(seq_lens.shape[0]):
+            # TODO: this loop process a sequence per iter, this is inefficient.
+            # Need optimize the performance later.
+
+            seq_len_q = 1
+            seq_len_kv = seq_lens[seq_idx]
+            end_q = start_q + seq_len_q
+            end_kv = start_kv + seq_len_kv
+
+            per_req_query = query[:, start_q:end_q, :]
+
+            # get key and value from cache. per_req_tokens contains the kv cache
+            # index for each token in the sequence.
+            req_pool_idx = req_pool_indices[seq_idx]
+            per_req_tokens = req_to_token[req_pool_idx, :seq_len_kv]
+            per_req_key = k_cache[per_req_tokens].movedim(0, query.dim() - 2)
+            per_req_value = v_cache[per_req_tokens].movedim(0, query.dim() - 2)
+
+            per_req_out = (
+                self.flex_attention(
+                    per_req_query.unsqueeze(0),
+                    per_req_key.unsqueeze(0),
+                    per_req_value.unsqueeze(0),
+                    block_mask=self.decode_block_masks[seq_idx],
+                    scale=scaling,
+                    enable_gqa=enable_gqa,
+                )
+                .squeeze(0)
+                .movedim(query.dim() - 2, 0)
+            )
+
+            output[start_q:end_q, :, :] = per_req_out
+            start_q, start_kv = end_q, end_kv
+
+        return output
+
+    def forward_extend(
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
+    ):
+        if layer.qk_head_dim != layer.v_head_dim:
+            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
+
+        use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
+
+        q_ = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+        o_ = o.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+
+        causal = True
+        if layer.is_cross_attention or layer.attn_type == AttentionType.ENCODER_ONLY:
+            raise NotImplementedError(
+                "TorchFlexAttnBackend does not support non-causal attention for now."
+            )
+
+        self._run_flex_forward_extend(
+            q_,
+            o_,
+            forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id),
+            forward_batch.token_to_kv_pool.get_value_buffer(layer.layer_id),
+            forward_batch.req_to_token_pool.req_to_token,
+            forward_batch.req_pool_indices,
+            forward_batch.seq_lens,
+            forward_batch.extend_prefix_lens,
+            forward_batch.extend_seq_lens,
+            scaling=layer.scaling,
+            enable_gqa=use_gqa,
+            causal=causal,
+        )
+        return o
+
+    def forward_decode(
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
+    ):
+        # During torch.compile, there is a bug in rotary_emb that causes the
+        # output value to have a 3D tensor shape. This reshapes the output correctly.
+        q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)
+
+        if layer.qk_head_dim != layer.v_head_dim:
+            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
+
+        use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
+        q_ = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+        o_ = o.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+
+        self._run_flex_forward_decode(
+            q_,
+            o_,
+            forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id),
+            forward_batch.token_to_kv_pool.get_value_buffer(layer.layer_id),
+            forward_batch.req_to_token_pool.req_to_token,
+            forward_batch.req_pool_indices,
+            forward_batch.seq_lens,
+            scaling=layer.scaling,
+            enable_gqa=use_gqa,
+            causal=False,
+        )
+
+        return o
+
+    def support_triton(self):
+        return False

sglang/srt/layers/attention/torch_native_backend.py

@@ -193,10 +193,13 @@ class TorchNativeAttnBackend(AttentionBackend):
         else:
             o = torch.empty_like(q)
 
+        if layer.is_cross_attention:
+            cache_loc = forward_batch.encoder_out_cache_loc
+        else:
+            cache_loc = forward_batch.out_cache_loc
+
         if save_kv_cache:
-            forward_batch.token_to_kv_pool.set_kv_buffer(
-                layer, forward_batch.out_cache_loc, k, v
-            )
+            forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
         use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 
@@ -241,10 +244,13 @@ class TorchNativeAttnBackend(AttentionBackend):
         else:
             o = torch.empty_like(q)
 
+        if layer.is_cross_attention:
+            cache_loc = forward_batch.encoder_out_cache_loc
+        else:
+            cache_loc = forward_batch.out_cache_loc
+
         if save_kv_cache:
-            forward_batch.token_to_kv_pool.set_kv_buffer(
-                layer, forward_batch.out_cache_loc, k, v
-            )
+            forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
 
         use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 

sglang/srt/layers/attention/triton_backend.py

@@ -12,12 +12,17 @@ from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_trito
 from sglang.srt.layers.dp_attention import get_attention_tp_size
 from sglang.srt.layers.radix_attention import AttentionType
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
-from sglang.srt.utils import get_bool_env_var, get_device_core_count, next_power_of_2
+from sglang.srt.utils import (
+    get_bool_env_var,
+    get_device_core_count,
+    get_int_env_var,
+    next_power_of_2,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
     from sglang.srt.model_executor.model_runner import ModelRunner
-    from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
+    from sglang.srt.speculative.spec_info import SpecInput
 
 
 def logit_capping_mod(logit_capping_method, logit_cap):
@@ -80,7 +85,13 @@ class TritonAttnBackend(AttentionBackend):
         self.num_kv_head = model_runner.model_config.get_num_kv_heads(
             get_attention_tp_size()
         )
-        self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(0).shape[-1]
+        if model_runner.is_hybrid_gdn:
+            # For hybrid linear models, layer_id = 0 may not be full attention
+            self.v_head_dim = model_runner.token_to_kv_pool.get_v_head_dim()
+        else:
+            self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(0).shape[
+                -1
+            ]
         self.max_context_len = model_runner.model_config.context_len
         self.device = model_runner.device
         self.device_core_count = get_device_core_count(model_runner.gpu_id)
@@ -89,6 +100,29 @@ class TritonAttnBackend(AttentionBackend):
         )
         self.max_kv_splits = model_runner.server_args.triton_attention_num_kv_splits
 
+        # Decide whether enable deterministic inference with batch-invariant operations
+        self.enable_deterministic = (
+            model_runner.server_args.enable_deterministic_inference
+        )
+
+        # Configure deterministic inference settings
+        if self.enable_deterministic:
+            # Use fixed split tile size for batch invariance
+            self.split_tile_size = get_int_env_var(
+                "SGLANG_TRITON_DECODE_SPLIT_TILE_SIZE", 256
+            )
+            # Set static_kv_splits to False to use deterministic logic instead
+            self.static_kv_splits = False
+        else:
+            self.split_tile_size = (
+                model_runner.server_args.triton_attention_split_tile_size
+            )
+
+        if self.split_tile_size is not None:
+            self.max_kv_splits = (
+                self.max_context_len + self.split_tile_size - 1
+            ) // self.split_tile_size
+
         # Check arguments
         assert not (
             model_runner.sliding_window_size is not None
@@ -143,10 +177,26 @@ class TritonAttnBackend(AttentionBackend):
             num_group * num_seq == num_token
         ), f"num_seq({num_seq}), num_token({num_token}), something goes wrong!"
 
-        if self.static_kv_splits or self.device_core_count <= 0:
+        # Legacy dynamic splitting logic (non-deterministic)
+        if (
+            self.static_kv_splits or self.device_core_count <= 0
+        ) and not self.enable_deterministic:
             num_kv_splits.fill_(self.max_kv_splits)
             return
 
+        # deterministic
+        if self.split_tile_size is not None and self.enable_deterministic:
+            # expand seq_lens to match num_token
+            if num_group > 1:
+                expanded_seq_lens = seq_lens.repeat_interleave(num_group)
+            else:
+                expanded_seq_lens = seq_lens
+
+            num_kv_splits[:] = (
+                expanded_seq_lens + self.split_tile_size - 1
+            ) // self.split_tile_size
+            return
+
         if num_seq < 256:
             SCHEDULE_SEQ = 256
         else:
@@ -432,7 +482,7 @@ class TritonAttnBackend(AttentionBackend):
         seq_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
     ):
         assert encoder_lens is None, "Not supported"
         window_kv_indptr = self.window_kv_indptr
@@ -588,7 +638,7 @@ class TritonAttnBackend(AttentionBackend):
         seq_lens_sum: int,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[Union[EagleDraftInput, EagleVerifyInput]],
+        spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
     ):
         # NOTE: encoder_lens expected to be zeros or None
@@ -833,7 +883,7 @@ class TritonMultiStepDraftBackend:
         topk: int,
         speculative_num_steps: int,
     ):
-        from sglang.srt.speculative.eagle_utils import generate_draft_decode_kv_indices
+        from sglang.srt.speculative.spec_utils import generate_draft_decode_kv_indices
 
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps

sglang/srt/layers/attention/trtllm_mha_backend.py

@@ -20,12 +20,10 @@ from sglang.srt.utils import is_flashinfer_available
 if is_flashinfer_available():
     import flashinfer
 
-from sglang.srt.speculative.eagle_utils import EagleDraftInput
-
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
     from sglang.srt.model_executor.model_runner import ModelRunner
-    from sglang.srt.speculative.spec_info import SpecInfo
+    from sglang.srt.speculative.spec_info import SpecInput
 
 # Constants
 DEFAULT_WORKSPACE_SIZE_MB = (
@@ -201,7 +199,7 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
         seq_lens: torch.Tensor,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
     ):
         """Initialize metadata for CUDA graph capture."""
         metadata = TRTLLMMHAMetadata()
@@ -314,7 +312,7 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
         seq_lens_sum: int,
         encoder_lens: Optional[torch.Tensor],
         forward_mode: ForwardMode,
-        spec_info: Optional[SpecInfo],
+        spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
     ):
         """Replay CUDA graph with new inputs."""
@@ -661,7 +659,7 @@ class TRTLLMHAAttnMultiStepDraftBackend(FlashInferMultiStepDraftBackend):
         forward_batch: ForwardBatch,
     ):
         assert forward_batch.spec_info is not None
-        assert isinstance(forward_batch.spec_info, EagleDraftInput)
+        assert forward_batch.spec_info.is_draft_input()
 
         for i in range(self.speculative_num_steps - 1):
             self.attn_backends[i].init_forward_metadata_capture_cuda_graph(
@@ -678,7 +676,7 @@ class TRTLLMHAAttnMultiStepDraftBackend(FlashInferMultiStepDraftBackend):
         self, forward_batch: ForwardBatch, bs: int
     ):
         assert forward_batch.spec_info is not None
-        assert isinstance(forward_batch.spec_info, EagleDraftInput)
+        assert forward_batch.spec_info.is_draft_input()
 
         for i in range(self.speculative_num_steps - 1):