sglang 0.2.11__py3-none-any.whl → 0.2.13__py3-none-any.whl

sglang/srt/model_executor/forward_batch_info.py

@@ -16,13 +16,17 @@ limitations under the License.
 """ModelRunner runs the forward passes of the models."""
 from dataclasses import dataclass
 from enum import IntEnum, auto
-from typing import List
+from typing import TYPE_CHECKING, List, Optional
 
 import numpy as np
 import torch
 
+from sglang.srt.managers.schedule_batch import ScheduleBatch
 from sglang.srt.mem_cache.memory_pool import BaseTokenToKVPool, ReqToTokenPool
 
+if TYPE_CHECKING:
+    from sglang.srt.model_executor.model_runner import ModelRunner
+
 
 class ForwardMode(IntEnum):
     # Prefill a new sequence. This is deprecated now. "EXTEND" covers this case.
@@ -39,25 +43,33 @@ class InputMetadata:
 
     forward_mode: ForwardMode
     batch_size: int
-    total_num_tokens: int
     req_pool_indices: torch.Tensor
     seq_lens: torch.Tensor
-    positions: torch.Tensor
     req_to_token_pool: ReqToTokenPool
     token_to_kv_pool: BaseTokenToKVPool
 
-    # For extend
-    extend_seq_lens: torch.Tensor
-    extend_start_loc: torch.Tensor
-    extend_no_prefix: bool
-
     # Output location of the KV cache
-    out_cache_loc: torch.Tensor = None
+    out_cache_loc: torch.Tensor
+
+    total_num_tokens: int = None
+
+    # Position information
+    positions: torch.Tensor = None
+
+    # For extend
+    extend_seq_lens: torch.Tensor = None
+    extend_start_loc: torch.Tensor = None
+    extend_no_prefix: bool = None
 
     # Output options
     return_logprob: bool = False
     top_logprobs_nums: List[int] = None
 
+    # For multimodal
+    pixel_values: List[torch.Tensor] = None
+    image_sizes: List[List[int]] = None
+    image_offsets: List[int] = None
+
     # Trition attention backend
     triton_max_seq_len: int = 0
     triton_max_extend_len: int = 0
@@ -70,107 +82,175 @@ class InputMetadata:
     flashinfer_decode_wrapper: "BatchDecodeWithPagedKVCacheWrapper" = None
     flashinfer_use_ragged: bool = False
 
-    @classmethod
-    def create(
-        cls,
-        model_runner,
-        forward_mode,
-        req_pool_indices,
-        seq_lens,
-        prefix_lens,
-        position_ids_offsets,
-        out_cache_loc,
-        top_logprobs_nums=None,
-        return_logprob=False,
-        skip_flashinfer_init=False,
-    ):
-        flashinfer_use_ragged = False
-        if not skip_flashinfer_init and not model_runner.server_args.disable_flashinfer:
-            if forward_mode != ForwardMode.DECODE and int(torch.sum(seq_lens)) > 4096:
-                flashinfer_use_ragged = True
-            init_flashinfer_args(
-                forward_mode,
-                model_runner,
-                req_pool_indices,
-                seq_lens,
-                prefix_lens,
-                model_runner.flashinfer_decode_wrapper,
-                flashinfer_use_ragged,
+    def init_multimuldal_info(self, batch: ScheduleBatch):
+        reqs = batch.reqs
+        self.pixel_values = [r.pixel_values for r in reqs]
+        self.image_sizes = [r.image_size for r in reqs]
+        self.image_offsets = [
+            (
+                (r.image_offset - len(r.prefix_indices))
+                if r.image_offset is not None
+                else 0
             )
+            for r in reqs
+        ]
 
-        batch_size = len(req_pool_indices)
+    def compute_positions(self, batch: ScheduleBatch):
+        position_ids_offsets = batch.position_ids_offsets
 
-        if forward_mode == ForwardMode.DECODE:
-            positions = ((seq_lens - 1) + position_ids_offsets).to(torch.int64)
-            extend_seq_lens = extend_start_loc = extend_no_prefix = None
-            if not model_runner.server_args.disable_flashinfer:
-                # This variable is not needed in this case,
-                # we do not compute it to make it compatbile with cuda graph.
-                total_num_tokens = None
+        if self.forward_mode == ForwardMode.DECODE:
+            if True:
+                self.positions = self.seq_lens - 1
             else:
-                total_num_tokens = int(torch.sum(seq_lens))
+                # Deprecated
+                self.positions = (self.seq_lens - 1) + position_ids_offsets
         else:
-            seq_lens_cpu = seq_lens.cpu().numpy()
-            prefix_lens_cpu = prefix_lens.cpu().numpy()
-            position_ids_offsets_cpu = position_ids_offsets.cpu().numpy()
-            positions = torch.tensor(
-                np.concatenate(
-                    [
-                        np.arange(
-                            prefix_lens_cpu[i] + position_ids_offsets_cpu[i],
-                            seq_lens_cpu[i] + position_ids_offsets_cpu[i],
-                        )
-                        for i in range(batch_size)
-                    ],
-                    axis=0,
-                ),
-                device="cuda",
-            )
-            extend_seq_lens = seq_lens - prefix_lens
-            extend_start_loc = torch.zeros_like(seq_lens)
-            extend_start_loc[1:] = torch.cumsum(extend_seq_lens[:-1], dim=0)
-            extend_no_prefix = torch.all(prefix_lens == 0)
-            total_num_tokens = int(torch.sum(seq_lens))
+            if True:
+                self.positions = torch.tensor(
+                    np.concatenate(
+                        [
+                            np.arange(len(req.prefix_indices), len(req.fill_ids))
+                            for req in batch.reqs
+                        ],
+                        axis=0,
+                    ),
+                    device="cuda",
+                )
+            else:
+                # Deprecated
+                position_ids_offsets_cpu = position_ids_offsets.cpu().numpy()
+                self.positions = torch.tensor(
+                    np.concatenate(
+                        [
+                            np.arange(
+                                len(req.prefix_indices) + position_ids_offsets_cpu[i],
+                                len(req.fill_ids) + position_ids_offsets_cpu[i],
+                            )
+                            for i, req in enumerate(batch.reqs)
+                        ],
+                        axis=0,
+                    ),
+                    device="cuda",
+                )
+
+        # Positions should be in long type
+        self.positions = self.positions.to(torch.int64)
+
+    def compute_extend_infos(self, batch: ScheduleBatch):
+        if self.forward_mode == ForwardMode.DECODE:
+            self.extend_seq_lens = self.extend_start_loc = self.extend_no_prefix = None
+        else:
+            extend_lens_cpu = [
+                len(r.fill_ids) - len(r.prefix_indices) for r in batch.reqs
+            ]
+            self.extend_seq_lens = torch.tensor(extend_lens_cpu, device="cuda")
+            self.extend_start_loc = torch.zeros_like(self.seq_lens)
+            self.extend_start_loc[1:] = torch.cumsum(self.extend_seq_lens[:-1], dim=0)
+            self.extend_no_prefix = all(len(r.prefix_indices) == 0 for r in batch.reqs)
 
+    @classmethod
+    def from_schedule_batch(
+        cls,
+        model_runner: "ModelRunner",
+        batch: ScheduleBatch,
+        forward_mode: ForwardMode,
+    ):
         ret = cls(
             forward_mode=forward_mode,
-            batch_size=batch_size,
-            total_num_tokens=total_num_tokens,
-            req_pool_indices=req_pool_indices,
-            seq_lens=seq_lens,
-            positions=positions,
+            batch_size=batch.batch_size(),
+            req_pool_indices=batch.req_pool_indices,
+            seq_lens=batch.seq_lens,
             req_to_token_pool=model_runner.req_to_token_pool,
             token_to_kv_pool=model_runner.token_to_kv_pool,
-            out_cache_loc=out_cache_loc,
-            extend_seq_lens=extend_seq_lens,
-            extend_start_loc=extend_start_loc,
-            extend_no_prefix=extend_no_prefix,
-            return_logprob=return_logprob,
-            top_logprobs_nums=top_logprobs_nums,
-            flashinfer_prefill_wrapper_ragged=model_runner.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=model_runner.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=model_runner.flashinfer_decode_wrapper,
-            flashinfer_use_ragged=flashinfer_use_ragged,
+            out_cache_loc=batch.out_cache_loc,
+            return_logprob=batch.return_logprob,
+            top_logprobs_nums=batch.top_logprobs_nums,
         )
 
+        ret.compute_positions(batch)
+
+        ret.compute_extend_infos(batch)
+
+        if (
+            forward_mode != ForwardMode.DECODE
+            or model_runner.server_args.disable_flashinfer
+        ):
+            ret.total_num_tokens = int(torch.sum(ret.seq_lens))
+
+        if forward_mode != ForwardMode.DECODE:
+            ret.init_multimuldal_info(batch)
+
+        prefix_lens = None
+        if forward_mode != ForwardMode.DECODE:
+            prefix_lens = torch.tensor(
+                [len(r.prefix_indices) for r in batch.reqs], device="cuda"
+            )
+
         if model_runner.server_args.disable_flashinfer:
-            (
-                ret.triton_max_seq_len,
-                ret.triton_max_extend_len,
-                ret.triton_start_loc,
-                ret.triton_prefix_lens,
-            ) = init_triton_args(forward_mode, seq_lens, prefix_lens)
+            ret.init_triton_args(batch, prefix_lens)
+
+        flashinfer_use_ragged = False
+        if not model_runner.server_args.disable_flashinfer:
+            if (
+                forward_mode != ForwardMode.DECODE
+                and int(torch.sum(ret.seq_lens)) > 4096
+                and model_runner.sliding_window_size is None
+            ):
+                flashinfer_use_ragged = True
+            ret.init_flashinfer_handlers(
+                model_runner, prefix_lens, flashinfer_use_ragged
+            )
 
         return ret
 
+    def init_triton_args(self, batch: ScheduleBatch, prefix_lens):
+        """Init auxiliary variables for triton attention backend."""
+        self.triton_max_seq_len = int(torch.max(self.seq_lens))
+        self.triton_prefix_lens = prefix_lens
+        self.triton_start_loc = torch.zeros_like(self.seq_lens, dtype=torch.int32)
+        self.triton_start_loc[1:] = torch.cumsum(self.seq_lens[:-1], dim=0)
+
+        if self.forward_mode == ForwardMode.DECODE:
+            self.triton_max_extend_len = None
+        else:
+            extend_seq_lens = self.seq_lens - prefix_lens
+            self.triton_max_extend_len = int(torch.max(extend_seq_lens))
+
+    def init_flashinfer_handlers(
+        self,
+        model_runner,
+        prefix_lens,
+        flashinfer_use_ragged,
+    ):
+        update_flashinfer_indices(
+            self.forward_mode,
+            model_runner,
+            self.req_pool_indices,
+            self.seq_lens,
+            prefix_lens,
+            flashinfer_use_ragged=flashinfer_use_ragged,
+        )
 
-def init_flashinfer_args(
+        (
+            self.flashinfer_prefill_wrapper_ragged,
+            self.flashinfer_prefill_wrapper_paged,
+            self.flashinfer_decode_wrapper,
+            self.flashinfer_use_ragged,
+        ) = (
+            model_runner.flashinfer_prefill_wrapper_ragged,
+            model_runner.flashinfer_prefill_wrapper_paged,
+            model_runner.flashinfer_decode_wrapper,
+            flashinfer_use_ragged,
+        )
+
+
+def update_flashinfer_indices(
     forward_mode,
     model_runner,
     req_pool_indices,
     seq_lens,
     prefix_lens,
-    flashinfer_decode_wrapper,
+    flashinfer_decode_wrapper=None,
     flashinfer_use_ragged=False,
 ):
     """Init auxiliary variables for FlashInfer attention backend."""
@@ -178,79 +258,136 @@ def init_flashinfer_args(
     num_kv_heads = model_runner.model_config.get_num_kv_heads(model_runner.tp_size)
     head_dim = model_runner.model_config.head_dim
     batch_size = len(req_pool_indices)
-    total_num_tokens = int(torch.sum(seq_lens))
-
-    if flashinfer_use_ragged:
-        paged_kernel_lens = prefix_lens
-    else:
-        paged_kernel_lens = seq_lens
-
-    kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
-    kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
-    req_pool_indices_cpu = req_pool_indices.cpu().numpy()
-    paged_kernel_lens_cpu = paged_kernel_lens.cpu().numpy()
-    kv_indices = torch.cat(
-        [
-            model_runner.req_to_token_pool.req_to_token[
-                req_pool_indices_cpu[i], : paged_kernel_lens_cpu[i]
-            ]
-            for i in range(batch_size)
-        ],
-        dim=0,
-    ).contiguous()
-    kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
-
-    if forward_mode == ForwardMode.DECODE:
-        flashinfer_decode_wrapper.end_forward()
-        flashinfer_decode_wrapper.begin_forward(
-            kv_indptr,
-            kv_indices,
-            kv_last_page_len,
-            num_qo_heads,
-            num_kv_heads,
-            head_dim,
-            1,
-        )
-    else:
-        # extend part
-        qo_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
-        qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0)
 
+    if model_runner.sliding_window_size is None:
         if flashinfer_use_ragged:
-            model_runner.flashinfer_prefill_wrapper_ragged.end_forward()
-            model_runner.flashinfer_prefill_wrapper_ragged.begin_forward(
-                qo_indptr,
+            paged_kernel_lens = prefix_lens
+        else:
+            paged_kernel_lens = seq_lens
+
+        kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
+        kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
+        req_pool_indices_cpu = req_pool_indices.cpu().numpy()
+        paged_kernel_lens_cpu = paged_kernel_lens.cpu().numpy()
+        kv_indices = torch.cat(
+            [
+                model_runner.req_to_token_pool.req_to_token[
+                    req_pool_indices_cpu[i], : paged_kernel_lens_cpu[i]
+                ]
+                for i in range(batch_size)
+            ],
+            dim=0,
+        ).contiguous()
+        kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
+
+        if forward_mode == ForwardMode.DECODE:
+            # CUDA graph uses different flashinfer_decode_wrapper
+            if flashinfer_decode_wrapper is None:
+                flashinfer_decode_wrapper = model_runner.flashinfer_decode_wrapper
+
+            flashinfer_decode_wrapper.end_forward()
+            flashinfer_decode_wrapper.begin_forward(
+                kv_indptr,
+                kv_indices,
+                kv_last_page_len,
+                num_qo_heads,
+                num_kv_heads,
+                head_dim,
+                1,
+            )
+        else:
+            # extend part
+            qo_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
+            qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0)
+
+            if flashinfer_use_ragged:
+                model_runner.flashinfer_prefill_wrapper_ragged.end_forward()
+                model_runner.flashinfer_prefill_wrapper_ragged.begin_forward(
+                    qo_indptr,
+                    qo_indptr,
+                    num_qo_heads,
+                    num_kv_heads,
+                    head_dim,
+                )
+
+            # cached part
+            model_runner.flashinfer_prefill_wrapper_paged.end_forward()
+            model_runner.flashinfer_prefill_wrapper_paged.begin_forward(
                 qo_indptr,
+                kv_indptr,
+                kv_indices,
+                kv_last_page_len,
                 num_qo_heads,
                 num_kv_heads,
                 head_dim,
+                1,
             )
-
-        # cached part
-        model_runner.flashinfer_prefill_wrapper_paged.end_forward()
-        model_runner.flashinfer_prefill_wrapper_paged.begin_forward(
-            qo_indptr,
-            kv_indptr,
-            kv_indices,
-            kv_last_page_len,
-            num_qo_heads,
-            num_kv_heads,
-            head_dim,
-            1,
-        )
-
-
-def init_triton_args(forward_mode, seq_lens, prefix_lens):
-    """Init auxiliary variables for triton attention backend."""
-    batch_size = len(seq_lens)
-    max_seq_len = int(torch.max(seq_lens))
-    start_loc = torch.zeros((batch_size,), dtype=torch.int32, device="cuda")
-    start_loc[1:] = torch.cumsum(seq_lens[:-1], dim=0)
-
-    if forward_mode == ForwardMode.DECODE:
-        max_extend_len = None
     else:
-        extend_seq_lens = seq_lens - prefix_lens
-        max_extend_len = int(torch.max(extend_seq_lens))
-
-    return max_seq_len, max_extend_len, start_loc, prefix_lens
+        # window attention use paged only
+        kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
+        for wrapper_id in range(2):
+            if wrapper_id == 0:
+                if forward_mode == ForwardMode.DECODE:
+                    paged_kernel_lens = torch.minimum(
+                        seq_lens, torch.tensor(model_runner.sliding_window_size + 1)
+                    )
+                else:
+                    paged_kernel_lens = torch.minimum(
+                        seq_lens,
+                        torch.tensor(model_runner.sliding_window_size)
+                        + seq_lens
+                        - prefix_lens,
+                    )
+            else:
+                paged_kernel_lens = seq_lens
+
+            kv_start_idx = seq_lens - paged_kernel_lens
+
+            kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
+            kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
+            req_pool_indices_cpu = req_pool_indices.cpu().numpy()
+            paged_kernel_lens_cpu = paged_kernel_lens.cpu().numpy()
+            kv_indices = torch.cat(
+                [
+                    model_runner.req_to_token_pool.req_to_token[
+                        req_pool_indices_cpu[i],
+                        kv_start_idx[i] : kv_start_idx[i] + paged_kernel_lens_cpu[i],
+                    ]
+                    for i in range(batch_size)
+                ],
+                dim=0,
+            ).contiguous()
+
+            if forward_mode == ForwardMode.DECODE:
+                # CUDA graph uses different flashinfer_decode_wrapper
+                if flashinfer_decode_wrapper is None:
+                    flashinfer_decode_wrapper = model_runner.flashinfer_decode_wrapper
+
+                flashinfer_decode_wrapper[wrapper_id].end_forward()
+                flashinfer_decode_wrapper[wrapper_id].begin_forward(
+                    kv_indptr,
+                    kv_indices,
+                    kv_last_page_len,
+                    num_qo_heads,
+                    num_kv_heads,
+                    head_dim,
+                    1,
+                )
+            else:
+                # extend part
+                qo_indptr = torch.zeros(
+                    (batch_size + 1,), dtype=torch.int32, device="cuda"
+                )
+                qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0)
+
+                model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].end_forward()
+                model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].begin_forward(
+                    qo_indptr,
+                    kv_indptr,
+                    kv_indices,
+                    kv_last_page_len,
+                    num_qo_heads,
+                    num_kv_heads,
+                    head_dim,
+                    1,
+                )