sglang 0.2.9__py3-none-any.whl → 0.2.10__py3-none-any.whl

sglang/bench_latency.py

@@ -1,13 +1,13 @@
 """
 Benchmark the latency of a given model. It accepts arguments similar to those of launch_server.py.
 
-# Usage (latency test):
+# Usage (latency test) with dummy weights:
 python -m sglang.bench_latency --model-path meta-llama/Meta-Llama-3-8B-Instruct --load-format dummy
 
 # Usage (correctness test):
 python -m sglang.bench_latency --model-path TinyLlama/TinyLlama-1.1B-Chat-v0.4 --correct
 
-### Reference output:
+### Reference output (of the correctness test above, can be gpu dependent):
 prefill logits (first half) tensor([[-10.0312,  -9.5000,   0.8936,  ...,  -4.9414,  -3.2402,  -3.3633],
         [-10.0312,  -9.5000,   0.8936,  ...,  -4.9414,  -3.2402,  -3.3633],
         [ -9.1875, -10.2500,   2.7109,  ...,  -4.3359,  -4.0664,  -4.1328]],
@@ -31,7 +31,9 @@ import dataclasses
 import logging
 import multiprocessing
 import time
+from typing import Tuple
 
+import jsonlines
 import numpy as np
 import torch
 import torch.distributed as dist
@@ -47,25 +49,34 @@ from sglang.srt.utils import suppress_other_loggers
 
 @dataclasses.dataclass
 class BenchArgs:
-    batch_size: int = 1
+    batch_size: Tuple[int] = (1,)
     input_len: int = 1024
     output_len: int = 4
+    result_filename: str = ""
     correctness_test: bool = False
     # This is only used for correctness test
     cut_len: int = 4
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
-        parser.add_argument("--batch-size", type=int, default=BenchArgs.batch_size)
+        parser.add_argument(
+            "--batch-size", type=int, nargs="+", default=BenchArgs.batch_size
+        )
         parser.add_argument("--input-len", type=int, default=BenchArgs.input_len)
         parser.add_argument("--output-len", type=int, default=BenchArgs.output_len)
+        parser.add_argument(
+            "--result-filename", type=str, default=BenchArgs.result_filename
+        )
         parser.add_argument("--correctness-test", action="store_true")
         parser.add_argument("--cut-len", type=int, default=BenchArgs.cut_len)
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
-        attrs = [attr.name for attr in dataclasses.fields(cls)]
-        return cls(**{attr: getattr(args, attr) for attr in attrs})
+        # use the default value's type to case the args into correct types.
+        attrs = [(attr.name, type(attr.default)) for attr in dataclasses.fields(cls)]
+        return cls(
+            **{attr: attr_type(getattr(args, attr)) for attr, attr_type in attrs}
+        )
 
 
 def load_model(server_args, tp_rank):
@@ -93,7 +104,7 @@ def load_model(server_args, tp_rank):
     return model_runner, tokenizer
 
 
-def prepare_inputs(bench_args, tokenizer):
+def prepare_inputs_for_correctness_test(bench_args, tokenizer):
     prompts = [
         "The capital of France is",
         "The capital of the United Kindom is",
@@ -119,7 +130,9 @@ def prepare_inputs(bench_args, tokenizer):
     return input_ids, reqs
 
 
-def prepare_extend_inputs(bench_args, input_ids, reqs, model_runner):
+def prepare_extend_inputs_for_correctness_test(
+    bench_args, input_ids, reqs, model_runner
+):
     for i in range(len(reqs)):
         req = reqs[i]
         req.input_ids += input_ids[i][bench_args.cut_len :]
@@ -129,8 +142,8 @@ def prepare_extend_inputs(bench_args, input_ids, reqs, model_runner):
     return reqs
 
 
-def prepare_synthetic_inputs(bench_args, tokenizer):
-    input_ids = np.ones((bench_args.batch_size, bench_args.input_len), dtype=np.int32)
+def prepare_synthetic_inputs_for_latency_test(batch_size, input_len):
+    input_ids = np.ones((batch_size, input_len), dtype=np.int32)
     sampling_params = SamplingParams(
         temperature=0,
         max_new_tokens=BenchArgs.output_len,
@@ -179,7 +192,7 @@ def correctness_test(
     model_runner, tokenizer = load_model(server_args, tp_rank)
 
     # Prepare inputs
-    input_ids, reqs = prepare_inputs(bench_args, tokenizer)
+    input_ids, reqs = prepare_inputs_for_correctness_test(bench_args, tokenizer)
 
     if bench_args.cut_len > 0:
         # Prefill
@@ -187,7 +200,9 @@ def correctness_test(
         rank_print("prefill logits (first half)", next_token_logits)
 
     # Prepare extend inputs
-    reqs = prepare_extend_inputs(bench_args, input_ids, reqs, model_runner)
+    reqs = prepare_extend_inputs_for_correctness_test(
+        bench_args, input_ids, reqs, model_runner
+    )
 
     # Extend
     next_token_ids, next_token_logits, batch = extend(reqs, model_runner)
@@ -205,6 +220,68 @@ def correctness_test(
         rank_print(tokenizer.decode(output_ids[i]))
 
 
+@torch.inference_mode()
+def latency_test_run_once(
+    model_runner, rank_print, reqs, batch_size, input_len, output_len
+):
+
+    # Clear the pools.
+    model_runner.req_to_token_pool.clear()
+    model_runner.token_to_kv_pool.clear()
+
+    measurement_results = {
+        "run_name": "before",
+        "batch_size": batch_size,
+        "input_len": input_len,
+        "output_len": output_len,
+    }
+
+    tot_latency = 0
+
+    # Prefill
+    torch.cuda.synchronize()
+    tic = time.time()
+    next_token_ids, _, batch = extend(reqs, model_runner)
+    torch.cuda.synchronize()
+    prefill_latency = time.time() - tic
+    tot_latency += prefill_latency
+    throughput = input_len * batch_size / prefill_latency
+    rank_print(
+        f"Prefill. latency: {prefill_latency:6.5f} s, throughput: {throughput:9.2f} token/s"
+    )
+    measurement_results["prefill_latency"] = prefill_latency
+    measurement_results["prefill_throughput"] = throughput
+
+    # Decode
+    for i in range(output_len):
+        torch.cuda.synchronize()
+        tic = time.time()
+        next_token_ids, _ = decode(next_token_ids, batch, model_runner)
+        torch.cuda.synchronize()
+        latency = time.time() - tic
+        tot_latency += latency
+        throughput = batch_size / latency
+        if i < 5:
+            rank_print(
+                f"Decode.  latency: {latency:6.5f} s, throughput: {throughput:9.2f} token/s"
+            )
+    avg_decode_latency = (tot_latency - prefill_latency) / output_len
+    avg_decode_throughput = batch_size / avg_decode_latency
+    rank_print(
+        f"Decode.  avg latency: {avg_decode_latency:6.5f} s, avg throughput: {avg_decode_throughput:9.2f} token/s"
+    )
+    measurement_results["avg_decode_latency"] = avg_decode_latency
+    measurement_results["avg_decode_throughput"] = avg_decode_throughput
+
+    throughput = (input_len + output_len) * batch_size / tot_latency
+    rank_print(
+        f"Total. latency: {tot_latency:6.3f} s, throughput: {throughput:9.2f} token/s"
+    )
+    measurement_results["total_latency"] = tot_latency
+    measurement_results["total_throughput"] = throughput
+    return measurement_results
+
+
 def latency_test(
     server_args,
     bench_args,
@@ -218,62 +295,36 @@ def latency_test(
         f"max_batch_size={model_runner.max_total_num_tokens // (bench_args.input_len + bench_args.output_len)}"
     )
 
-    # Prepare inputs
-    reqs = prepare_synthetic_inputs(bench_args, tokenizer)
-
-    def clear():
-        model_runner.req_to_token_pool.clear()
-        model_runner.token_to_kv_pool.clear()
-
-    @torch.inference_mode()
-    def run_once(output_len):
-        # Prefill
-        torch.cuda.synchronize()
-        tot_latency = 0
-        tic = time.time()
-        next_token_ids, _, batch = extend(reqs, model_runner)
-        torch.cuda.synchronize()
-        prefill_latency = time.time() - tic
-        tot_latency += prefill_latency
-        throughput = bench_args.input_len * bench_args.batch_size / prefill_latency
-        rank_print(
-            f"Prefill. latency: {prefill_latency:6.5f} s, throughput: {throughput:9.2f} token/s"
-        )
+    # To make this PR easier to review, for now, only do the first element in batch_size tuple.
+    bench_args.batch_size = bench_args.batch_size[0]
 
-        # Decode
-        for i in range(output_len):
-            torch.cuda.synchronize()
-            tic = time.time()
-            next_token_ids, _ = decode(next_token_ids, batch, model_runner)
-            torch.cuda.synchronize()
-            latency = time.time() - tic
-            tot_latency += latency
-            throughput = bench_args.batch_size / latency
-            if i < 5:
-                rank_print(
-                    f"Decode.  latency: {latency:6.5f} s, throughput: {throughput:9.2f} token/s"
-                )
-        avg_decode_latency = (tot_latency - prefill_latency) / output_len
-        avg_decode_throughput = bench_args.batch_size / avg_decode_latency
-        rank_print(
-            f"Decode.  avg latency: {avg_decode_latency:6.5f} s, avg throughput: {avg_decode_throughput:9.2f} token/s"
-        )
-
-        throughput = (
-            (bench_args.input_len + bench_args.output_len)
-            * bench_args.batch_size
-            / tot_latency
-        )
-        rank_print(
-            f"Total. latency: {tot_latency:6.3f} s, throughput: {throughput:9.2f} token/s"
-        )
+    # Prepare inputs
+    reqs = prepare_synthetic_inputs_for_latency_test(
+        bench_args.batch_size, bench_args.input_len
+    )
 
     # Warm up
-    run_once(4)
-    clear()
+    latency_test_run_once(
+        model_runner, rank_print, reqs, bench_args.batch_size, bench_args.input_len, 4
+    )
 
     # Run again
-    run_once(bench_args.output_len)
+    result_list = []
+    result_list.append(
+        latency_test_run_once(
+            model_runner,
+            rank_print,
+            reqs,
+            bench_args.batch_size,
+            bench_args.input_len,
+            bench_args.output_len,
+        )
+    )
+
+    # Write results in jsonlines format.
+    if bench_args.result_filename:
+        with jsonlines.open(bench_args.result_filename, "a") as f:
+            f.write_all(result_list)
 
 
 def main(server_args, bench_args):

sglang/check_env.py

@@ -13,6 +13,7 @@ import torch
 PACKAGE_LIST = [
     "sglang",
     "flashinfer",
+    "triton",
     "requests",
     "tqdm",
     "numpy",
@@ -30,6 +31,7 @@ PACKAGE_LIST = [
     "zmq",
     "vllm",
     "outlines",
+    "multipart",
     "openai",
     "tiktoken",
     "anthropic",

sglang/lang/backend/runtime_endpoint.py

@@ -15,7 +15,6 @@ class RuntimeEndpoint(BaseBackend):
     def __init__(
         self,
         base_url: str,
-        auth_token: Optional[str] = None,
         api_key: Optional[str] = None,
         verify: Optional[str] = None,
     ):
@@ -23,13 +22,11 @@ class RuntimeEndpoint(BaseBackend):
         self.support_concate_and_append = True
 
         self.base_url = base_url
-        self.auth_token = auth_token
         self.api_key = api_key
         self.verify = verify
 
         res = http_request(
             self.base_url + "/get_model_info",
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -67,7 +64,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/generate",
             json={"text": prefix_str, "sampling_params": {"max_new_tokens": 0}},
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -79,7 +75,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/generate",
             json=data,
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -91,7 +86,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/generate",
             json=data,
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -139,7 +133,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/generate",
             json=data,
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -193,7 +186,6 @@ class RuntimeEndpoint(BaseBackend):
             self.base_url + "/generate",
             json=data,
             stream=True,
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -225,7 +217,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/generate",
             json=data,
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -243,7 +234,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/generate",
             json=data,
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )
@@ -267,7 +257,6 @@ class RuntimeEndpoint(BaseBackend):
         res = http_request(
             self.base_url + "/concate_and_append_request",
             json={"src_rids": src_rids, "dst_rid": dst_rid},
-            auth_token=self.auth_token,
             api_key=self.api_key,
             verify=self.verify,
         )

sglang/srt/hf_transformers_utils.py

@@ -19,7 +19,7 @@ import functools
 import json
 import os
 import warnings
-from typing import AbstractSet, Collection, Dict, Literal, Optional, Type, Union
+from typing import AbstractSet, Collection, Dict, List, Literal, Optional, Type, Union
 
 from huggingface_hub import snapshot_download
 from transformers import (
@@ -259,7 +259,7 @@ class TiktokenTokenizer:
                 Literal["all"], AbstractSet[str]
             ] = set(),  # noqa: B006
             disallowed_special: Union[Literal["all"], Collection[str]] = "all",
-        ) -> list[int]:
+        ) -> List[int]:
             if isinstance(allowed_special, set):
                 allowed_special |= self._default_allowed_special
             return tiktoken.Encoding.encode(

sglang/srt/layers/extend_attention.py

@@ -57,6 +57,8 @@ def _fwd_kernel(
     stride_buf_vh,
     stride_req_to_tokens_b,
     BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
+    BLOCK_DV: tl.constexpr,
     BLOCK_M: tl.constexpr,
     BLOCK_N: tl.constexpr,
     logit_cap: tl.constexpr,
@@ -75,8 +77,10 @@ def _fwd_kernel(
     cur_batch_req_idx = tl.load(B_req_idx + cur_seq)
 
     offs_d = tl.arange(0, BLOCK_DMODEL)
+    offs_dv = tl.arange(0, BLOCK_DV)
     offs_m = tl.arange(0, BLOCK_M)
     mask_m = (cur_block_m * BLOCK_M + offs_m) < cur_seq_len_extend
+
     offs_q = (
         (cur_seq_extend_start_contiguous + cur_block_m * BLOCK_M + offs_m[:, None])
         * stride_qbs
@@ -85,10 +89,20 @@ def _fwd_kernel(
     )
     q = tl.load(Q_Extend + offs_q, mask=mask_m[:, None], other=0.0)
 
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        offs_qpe = (
+            (cur_seq_extend_start_contiguous + cur_block_m * BLOCK_M + offs_m[:, None])
+            * stride_qbs
+            + cur_head * stride_qh
+            + offs_dpe[None, :]
+        )
+        qpe = tl.load(Q_Extend + offs_qpe, mask=mask_m[:, None], other=0.0)
+
     # stage1: compute scores with prefix
     offs_n = tl.arange(0, BLOCK_N)
 
-    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DV], dtype=tl.float32)
     deno = tl.zeros([BLOCK_M], dtype=tl.float32)
     e_max = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
 
@@ -110,6 +124,18 @@ def _fwd_kernel(
 
         qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
         qk += tl.dot(q, k)
+        if BLOCK_DPE > 0:
+            offs_kpe = (
+                offs_kv_loc[None, :] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_dpe[:, None]
+            )
+            kpe = tl.load(
+                K_Buffer + offs_kpe,
+                mask=mask_n[None, :],
+                other=0.0,
+            )
+            qk += tl.dot(qpe, kpe)
         qk *= sm_scale
 
         if logit_cap > 0:
@@ -125,7 +151,7 @@ def _fwd_kernel(
         offs_buf_v = (
             offs_kv_loc[:, None] * stride_buf_vbs
             + cur_kv_head * stride_buf_vh
-            + offs_d[None, :]
+            + offs_dv[None, :]
         )
         v = tl.load(V_Buffer + offs_buf_v, mask=mask_n[:, None], other=0.0)
         p = p.to(v.dtype)
@@ -150,6 +176,21 @@ def _fwd_kernel(
 
         qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
         qk += tl.dot(q, k)
+
+        if BLOCK_DPE > 0:
+            offs_kpe = (
+                (cur_seq_extend_start_contiguous + start_n + offs_n[None, :])
+                * stride_kbs
+                + cur_kv_head * stride_kh
+                + offs_dpe[:, None]
+            )
+            kpe = tl.load(
+                K_Extend + offs_kpe,
+                mask=mask_n[None, :],
+                other=0.0,
+            )
+            qk += tl.dot(qpe, kpe)
+
         qk *= sm_scale
 
         if logit_cap > 0:
@@ -169,7 +210,7 @@ def _fwd_kernel(
         offs_v = (
             (cur_seq_extend_start_contiguous + start_n + offs_n[:, None]) * stride_vbs
             + cur_kv_head * stride_vh
-            + offs_d[None, :]
+            + offs_dv[None, :]
         )
         v = tl.load(V_Extend + offs_v, mask=mask_n[:, None], other=0.0)
         p = p.to(v.dtype)
@@ -181,7 +222,7 @@ def _fwd_kernel(
         (cur_seq_extend_start_contiguous + cur_block_m * BLOCK_M + offs_m[:, None])
         * stride_obs
         + cur_head * stride_oh
-        + offs_d[None, :]
+        + offs_dv[None, :]
     )
     tl.store(O_Extend + offs_o, acc / deno[:, None], mask=mask_m[:, None])
 
@@ -217,8 +258,17 @@ def extend_attention_fwd(
         o_extend.shape[-1],
     )
 
-    assert Lq == Lk and Lk == Lv and Lv == Lo
-    assert Lq in {16, 32, 64, 128, 256}
+    assert Lq == Lk and Lv == Lo
+    assert Lq in {16, 32, 64, 128, 256, 576}
+    assert Lv in {16, 32, 64, 128, 256, 512}
+
+    if Lq == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    else:
+        BLOCK_DMODEL = Lq
+        BLOCK_DPE = 0
+    BLOCK_DV = Lv
 
     if CUDA_CAPABILITY[0] >= 8:
         BLOCK_M, BLOCK_N = (128, 128) if Lq <= 128 else (64, 64)
@@ -260,7 +310,9 @@ def extend_attention_fwd(
         v_buffer.stride(0),
         v_buffer.stride(1),
         req_to_tokens.stride(0),
-        BLOCK_DMODEL=Lq,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
+        BLOCK_DV=BLOCK_DV,
         BLOCK_M=BLOCK_M,
         BLOCK_N=BLOCK_N,
         num_warps=num_warps,

sglang/srt/layers/radix_attention.py

@@ -38,16 +38,22 @@ class RadixAttention(nn.Module):
         num_kv_heads: int,
         layer_id: int,
         logit_cap: int = -1,
+        v_head_dim: int = -1,
     ):
         super().__init__()
         self.tp_q_head_num = num_heads
         self.tp_k_head_num = num_kv_heads
         self.tp_v_head_num = num_kv_heads
         self.head_dim = head_dim
+        self.qk_head_dim = head_dim
+        self.v_head_dim = v_head_dim if v_head_dim != -1 else head_dim
         self.scaling = scaling
         self.layer_id = layer_id
 
-        if not global_server_args_dict.get("disable_flashinfer", False):
+        if (
+            not global_server_args_dict.get("disable_flashinfer", False)
+            and self.qk_head_dim == self.v_head_dim
+        ):
             self.extend_forward = self.extend_forward_flashinfer
             self.decode_forward = self.decode_forward_flashinfer
         else:
@@ -57,13 +63,17 @@ class RadixAttention(nn.Module):
         self.logit_cap = logit_cap if logit_cap is not None and logit_cap > 0 else 0
 
     def extend_forward_triton(self, q, k, v, input_metadata: InputMetadata):
-        o = torch.empty_like(q)
+        if self.qk_head_dim != self.v_head_dim:
+            o = q.new_empty((q.shape[0], self.tp_q_head_num * self.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
         self.store_kv_cache(k, v, input_metadata)
         extend_attention_fwd(
-            q.view(-1, self.tp_q_head_num, self.head_dim),
+            q.view(-1, self.tp_q_head_num, self.qk_head_dim),
             k.contiguous(),
             v.contiguous(),
-            o.view(-1, self.tp_q_head_num, self.head_dim),
+            o.view(-1, self.tp_q_head_num, self.v_head_dim),
             input_metadata.token_to_kv_pool.get_key_buffer(self.layer_id),
             input_metadata.token_to_kv_pool.get_value_buffer(self.layer_id),
             input_metadata.req_to_token_pool.req_to_token,
@@ -82,14 +92,17 @@ class RadixAttention(nn.Module):
         return o
 
     def decode_forward_triton(self, q, k, v, input_metadata: InputMetadata):
-        o = torch.empty_like(q)
+        if self.qk_head_dim != self.v_head_dim:
+            o = q.new_empty((q.shape[0], self.tp_q_head_num * self.v_head_dim))
+        else:
+            o = torch.empty_like(q)
         self.store_kv_cache(k, v, input_metadata)
 
         token_attention_fwd(
-            q.view(-1, self.tp_q_head_num, self.head_dim),
+            q.view(-1, self.tp_q_head_num, self.qk_head_dim),
             input_metadata.token_to_kv_pool.get_key_buffer(self.layer_id),
             input_metadata.token_to_kv_pool.get_value_buffer(self.layer_id),
-            o.view(-1, self.tp_q_head_num, self.head_dim),
+            o.view(-1, self.tp_q_head_num, self.v_head_dim),
             input_metadata.req_to_token_pool.req_to_token,
             input_metadata.req_pool_indices,
             input_metadata.triton_start_loc,
@@ -160,8 +173,8 @@ class RadixAttention(nn.Module):
         return o.view(-1, self.tp_q_head_num * self.head_dim)
 
     def forward(self, q, k, v, input_metadata: InputMetadata):
-        k = k.view(-1, self.tp_k_head_num, self.head_dim)
-        v = v.view(-1, self.tp_v_head_num, self.head_dim)
+        k = k.view(-1, self.tp_k_head_num, self.qk_head_dim)
+        v = v.view(-1, self.tp_v_head_num, self.v_head_dim)
 
         if input_metadata.forward_mode == ForwardMode.EXTEND:
             return self.extend_forward(q, k, v, input_metadata)

sglang/srt/layers/token_attention.py

@@ -54,6 +54,7 @@ def _fwd_kernel_stage1(
     att_stride_h,
     kv_group_num: tl.constexpr,
     BLOCK_DMODEL: tl.constexpr,
+    BLOCK_DPE: tl.constexpr,
     BLOCK_N: tl.constexpr,
     logit_cap: tl.constexpr,
 ):
@@ -73,6 +74,10 @@ def _fwd_kernel_stage1(
 
     off_q = cur_batch * stride_qbs + cur_head * stride_qh + offs_d
 
+    if BLOCK_DPE > 0:
+        offs_dpe = BLOCK_DMODEL + tl.arange(0, BLOCK_DPE)
+        off_qpe = cur_batch * stride_qbs + cur_head * stride_qh + offs_dpe
+
     offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
 
     block_stard_index = start_n * BLOCK_N
@@ -97,6 +102,19 @@ def _fwd_kernel_stage1(
             other=0.0,
         ).to(REDUCE_TRITON_TYPE)
         att_value = tl.sum(q[None, :] * k, 1)
+        if BLOCK_DPE > 0:
+            qpe = tl.load(Q + off_qpe + start_mark).to(REDUCE_TRITON_TYPE)
+            offs_buf_kpe = (
+                k_loc[:, None] * stride_buf_kbs
+                + cur_kv_head * stride_buf_kh
+                + offs_dpe[None, :]
+            )
+            kpe = tl.load(
+                K_Buffer + offs_buf_kpe,
+                mask=offs_n_new[:, None] < cur_batch_end_index,
+                other=0.0,
+            ).to(REDUCE_TRITON_TYPE)
+            att_value += tl.sum(qpe[None, :] * kpe, 1)
         att_value *= sm_scale
 
         if logit_cap > 0:
@@ -192,7 +210,14 @@ def _token_att_m_fwd(
     # shape constraints
     Lq, Lk = q.shape[-1], k_buffer.shape[-1]
     assert Lq == Lk
-    assert Lk in {16, 32, 64, 128, 256}
+    assert Lk in {16, 32, 64, 128, 256, 576}
+
+    if Lk == 576:
+        BLOCK_DMODEL = 512
+        BLOCK_DPE = 64
+    else:
+        BLOCK_DMODEL = Lk
+        BLOCK_DPE = 0
 
     batch, head_num = B_req_idx.shape[0], q.shape[1]
 
@@ -220,7 +245,8 @@ def _token_att_m_fwd(
         k_buffer.stride(1),
         att_out.stride(0),
         kv_group_num=kv_group_num,
-        BLOCK_DMODEL=Lk,
+        BLOCK_DMODEL=BLOCK_DMODEL,
+        BLOCK_DPE=BLOCK_DPE,
         BLOCK_N=BLOCK,
         logit_cap=logit_cap,
         num_warps=num_warps,