更新README.md，修改Docker镜像版本，调整环境变量格式，优化基准测试脚本，添加新的基准测试功能，删除不再使用的示例文件。

benchmarks/kernels/benchmark_rmsnorm.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import itertools
+from typing import Optional, Union
+
+import torch
+from flashinfer.norm import fused_add_rmsnorm, rmsnorm
+from torch import nn
+
+from vllm import _custom_ops as vllm_ops
+from vllm.triton_utils import triton
+
+
+class HuggingFaceRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
+        orig_dtype = x.dtype
+        x = x.to(torch.float32)
+        if residual is not None:
+            x = x + residual.to(torch.float32)
+            residual = x.to(orig_dtype)
+
+        variance = x.pow(2).mean(dim=-1, keepdim=True)
+        x = x * torch.rsqrt(variance + self.variance_epsilon)
+        x = x.to(orig_dtype) * self.weight
+        if residual is None:
+            return x
+        else:
+            return x, residual
+
+
+def rmsnorm_naive(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+    eps: float = 1e-6,
+):
+    naive_norm = HuggingFaceRMSNorm(x.shape[-1], eps=eps)
+    naive_norm.weight = nn.Parameter(weight)
+    naive_norm = naive_norm.to(x.device)
+
+    orig_shape = x.shape
+    x = x.view(-1, x.shape[-1])
+    if residual is not None:
+        residual = residual.view(-1, residual.shape[-1])
+
+    output = naive_norm(x, residual)
+
+    if isinstance(output, tuple):
+        output = (output[0].view(orig_shape), output[1].view(orig_shape))
+    else:
+        output = output.view(orig_shape)
+    return output
+
+
+def rmsnorm_flashinfer(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+    eps: float = 1e-6,
+):
+    orig_shape = x.shape
+    x = x.view(-1, x.shape[-1])
+    if residual is not None:
+        residual = residual.view(-1, residual.shape[-1])
+
+    if residual is not None:
+        fused_add_rmsnorm(x, residual, weight, eps)
+        output = (x, residual)
+    else:
+        output = rmsnorm(x, weight, eps)
+
+    if isinstance(output, tuple):
+        output = (output[0].view(orig_shape), output[1].view(orig_shape))
+    else:
+        output = output.view(orig_shape)
+    return output
+
+
+def rmsnorm_vllm(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+    eps: float = 1e-6,
+):
+    orig_shape = x.shape
+    x = x.view(-1, x.shape[-1])
+    if residual is not None:
+        residual = residual.view(-1, residual.shape[-1])
+
+    if residual is not None:
+        vllm_ops.fused_add_rms_norm(x, residual, weight, eps)
+        output = (x, residual)
+    else:
+        out = torch.empty_like(x)
+        vllm_ops.rms_norm(out, x, weight, eps)
+        output = out
+
+    if isinstance(output, tuple):
+        output = (output[0].view(orig_shape), output[1].view(orig_shape))
+    else:
+        output = output.view(orig_shape)
+    return output
+
+
+def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):
+    dtype = torch.bfloat16
+    x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda")
+    weight = torch.ones(hidden_size, dtype=dtype, device="cuda")
+    residual = torch.randn_like(x) if use_residual else None
+
+    output_naive = rmsnorm_naive(
+        x.clone(), weight, residual.clone() if residual is not None else None
+    )
+    output_flashinfer = rmsnorm_flashinfer(
+        x.clone(), weight, residual.clone() if residual is not None else None
+    )
+    output_vllm = rmsnorm_vllm(
+        x.clone(), weight, residual.clone() if residual is not None else None
+    )
+
+    if use_residual:
+        output_naive = output_naive[0]
+        output_flashinfer = output_flashinfer[0]
+        output_vllm = output_vllm[0]
+
+    print(f"Naive output={output_naive}")
+    print(f"FlashInfer output={output_flashinfer}")
+    print(f"vLLM output={output_vllm}")
+
+    if torch.allclose(
+        output_naive, output_flashinfer, atol=1e-2, rtol=1e-2
+    ) and torch.allclose(output_naive, output_vllm, atol=1e-2, rtol=1e-2):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+
+
+batch_size_range = [2**i for i in range(0, 7, 2)]
+seq_length_range = [2**i for i in range(6, 11, 1)]
+head_num_range = [32, 48]
+configs = list(itertools.product(head_num_range, batch_size_range, seq_length_range))
+
+
+def get_benchmark(use_residual):
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["head_num", "batch_size", "seq_len"],
+            x_vals=[list(_) for _ in configs],
+            line_arg="provider",
+            line_vals=["huggingface", "flashinfer", "vllm"],
+            line_names=["HuggingFace", "FlashInfer", "vLLM"],
+            styles=[("blue", "-"), ("green", "-"), ("red", "-")],
+            ylabel="us",
+            plot_name=f"rmsnorm-perf-{'with' if use_residual else 'without'}-residual",
+            args={},
+        )
+    )
+    def benchmark(head_num, batch_size, seq_len, provider):
+        dtype = torch.bfloat16
+        hidden_size = head_num * 128  # assuming head_dim = 128
+
+        x = torch.randn(batch_size, seq_len, hidden_size, dtype=dtype, device="cuda")
+        weight = torch.ones(hidden_size, dtype=dtype, device="cuda")
+        residual = torch.randn_like(x) if use_residual else None
+
+        quantiles = [0.5, 0.2, 0.8]
+
+        if provider == "huggingface":
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: rmsnorm_naive(
+                    x.clone(),
+                    weight,
+                    residual.clone() if residual is not None else None,
+                ),
+                quantiles=quantiles,
+            )
+        elif provider == "flashinfer":
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: rmsnorm_flashinfer(
+                    x.clone(),
+                    weight,
+                    residual.clone() if residual is not None else None,
+                ),
+                quantiles=quantiles,
+            )
+        else:
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: rmsnorm_vllm(
+                    x.clone(),
+                    weight,
+                    residual.clone() if residual is not None else None,
+                ),
+                quantiles=quantiles,
+            )
+
+        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+    return benchmark
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=4,
+        help="Batch size",
+    )
+    parser.add_argument(
+        "--seq-len",
+        type=int,
+        default=128,
+        help="Sequence length",
+    )
+    parser.add_argument(
+        "--hidden-size",
+        type=int,
+        default=4096,
+        help="Hidden size (2nd dimension) of the sequence",
+    )
+    parser.add_argument(
+        "--use-residual", action="store_true", help="Whether to use residual connection"
+    )
+    parser.add_argument(
+        "--save-path",
+        type=str,
+        default="./configs/rmsnorm/",
+        help="Path to save rmsnorm benchmark results",
+    )
+
+    args = parser.parse_args()
+
+    # Run correctness test
+    calculate_diff(
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        hidden_size=args.hidden_size,
+        use_residual=args.use_residual,
+    )
+
+    # Get the benchmark function with proper use_residual setting
+    benchmark = get_benchmark(args.use_residual)
+    # Run performance benchmark
+    benchmark.run(print_data=True, save_path=args.save_path)

benchmarks/kernels/benchmark_rope.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 from itertools import accumulate
-from typing import List, Optional
+from typing import Optional
 
 import nvtx
 import torch
 
-from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
-                                                         get_rope)
-from vllm.utils import FlexibleArgumentParser, seed_everything
+from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
 
 
 def benchmark_rope_kernels_multi_lora(
@@ -20,49 +23,58 @@ def benchmark_rope_kernels_multi_lora(
     seed: int,
     device: str,
     max_position: int = 8192,
-    base: int = 10000,
+    base: float = 10000,
 ) -> None:
-    seed_everything(seed)
+    current_platform.seed_everything(seed)
     torch.set_default_device(device)
     if rotary_dim is None:
         rotary_dim = head_size
     # silulating serving 4 LoRAs
     scaling_factors = [1, 2, 4, 8]
     # batched RoPE can take multiple scaling factors
-    batched_rope = get_rope(head_size, rotary_dim, max_position, base,
-                            is_neox_style, {
-                                "type": "linear",
-                                "factor": tuple(scaling_factors)
-                            })
+    batched_rope = get_rope(
+        head_size,
+        rotary_dim,
+        max_position,
+        base,
+        is_neox_style,
+        {"rope_type": "linear", "factor": tuple(scaling_factors)},
+    )
     # non-batched RoPE takes only one scaling factor, we create multiple
     # instances to simulate the same behavior
-    non_batched_ropes: List[RotaryEmbedding] = []
+    non_batched_ropes: list[RotaryEmbedding] = []
     for scaling_factor in scaling_factors:
         non_batched_ropes.append(
-            get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
-                     {
-                         "type": "linear",
-                         "factor": (scaling_factor, )
-                     }))
+            get_rope(
+                head_size,
+                rotary_dim,
+                max_position,
+                base,
+                is_neox_style,
+                {"rope_type": "linear", "factor": (scaling_factor,)},
+            )
+        )
 
     positions = torch.randint(0, max_position, (batch_size, seq_len))
-    query = torch.randn(batch_size,
-                        seq_len,
-                        num_heads * head_size,
-                        dtype=dtype)
+    query = torch.randn(batch_size, seq_len, num_heads * head_size, dtype=dtype)
     key = torch.randn_like(query)
 
     # create query offsets for batched RoPE, we concat multiple kv cache
     # together and each query needs to find the right kv cache of its type
     offset_map = torch.tensor(
         list(
-            accumulate([0] + [
-                max_position * scaling_factor * 2
-                for scaling_factor in scaling_factors[:-1]
-            ])))
-    query_types = torch.randint(0,
-                                len(scaling_factors), (batch_size, seq_len),
-                                device=device)
+            accumulate(
+                [0]
+                + [
+                    max_position * scaling_factor * 2
+                    for scaling_factor in scaling_factors[:-1]
+                ]
+            )
+        )
+    )
+    query_types = torch.randint(
+        0, len(scaling_factors), (batch_size, seq_len), device=device
+    )
     # map query types to offsets
     query_offsets = offset_map[query_types]
     # the kernel takes flattened offsets
@@ -83,27 +95,28 @@ def benchmark_rope_kernels_multi_lora(
     torch.cuda.synchronize()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description="Benchmark the rotary embedding kernels.")
+        description="Benchmark the rotary embedding kernels."
+    )
     parser.add_argument("--is-neox-style", type=bool, default=True)
     parser.add_argument("--batch-size", type=int, default=16)
     parser.add_argument("--seq-len", type=int, default=512)
     parser.add_argument("--num-heads", type=int, default=8)
-    parser.add_argument("--head-size",
-                        type=int,
-                        choices=[64, 80, 96, 112, 120, 128, 192, 256],
-                        default=128)
+    parser.add_argument(
+        "--head-size",
+        type=int,
+        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+        default=128,
+    )
     parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32)
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["bfloat16", "float"],
-                        default="float")
+    parser.add_argument(
+        "--dtype", type=str, choices=["bfloat16", "float"], default="float"
+    )
     parser.add_argument("--seed", type=int, default=0)
-    parser.add_argument("--device",
-                        type=str,
-                        choices=["cuda:0", "cuda:1"],
-                        default="cuda:0")
+    parser.add_argument(
+        "--device", type=str, choices=["cuda:0", "cuda:1"], default="cuda:0"
+    )
     args = parser.parse_args()
     print(args)
 

benchmarks/kernels/benchmark_shapes.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 WEIGHT_SHAPES = {
     "ideal": [[4 * 256 * 32, 256 * 32]],
     "mistralai/Mistral-7B-v0.1/TP1": [
@@ -73,3 +76,19 @@ WEIGHT_SHAPES = {
         [7168, 8192],
     ],
 }
+
+WEIGHT_SHAPES_MOE = {
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
+        [8, 2, 4096, 28672],
+        [8, 2, 14336, 4096],
+    ],
+    "nm-testing/deepseekv2-lite": [
+        [64, 6, 2048, 1408],
+    ],
+    "ibm-granite/granite-3.0-1b-a400m": [
+        [32, 8, 1024, 1024],
+    ],
+    "ibm-granite/granite-3.0-3b-a800m": [
+        [40, 8, 1024, 1536],
+    ],
+}

benchmarks/kernels/benchmark_w8a8_block_fp8.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from sglang quantization/tuning_block_wise_kernel.py
+
+import argparse
+import json
+import multiprocessing as mp
+import os
+import time
+from datetime import datetime
+from typing import Any
+
+import torch
+import tqdm
+import triton
+
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    _w8a8_block_fp8_matmul,
+)
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+mp.set_start_method("spawn", force=True)
+
+assert current_platform.is_cuda(), (
+    "Only support tune w8a8 block fp8 kernel on CUDA device."
+)
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "float16": torch.float16,
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def w8a8_block_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: list[int],
+    config: dict[str, Any],
+    output_dtype: torch.dtype = torch.float16,
+) -> torch.Tensor:
+    """This function performs matrix multiplication with
+    block-wise quantization.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+
+    Args:
+        A: The input tensor, e.g., activation.
+        B: The input tensor, e.g., weight.
+        As: The per-token-group quantization scale for `A`.
+        Bs: The per-block quantization scale for `B`.
+        block_size: The block size for per-block quantization.
+                    It should be 2-dim, e.g., [128, 128].
+        output_dytpe: The dtype of the returned tensor.
+
+    Returns:
+        torch.Tensor: The result of matmul.
+    """
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+
+    assert A.shape[-1] == B.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
+    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    M = A.numel() // A.shape[-1]
+
+    assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
+    N, K = B.shape
+    assert triton.cdiv(N, block_n) == Bs.shape[0]
+    assert triton.cdiv(K, block_k) == Bs.shape[1]
+
+    C_shape = A.shape[:-1] + (N,)
+    C = A.new_empty(C_shape, dtype=output_dtype)
+
+    def grid(META):
+        return (
+            triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
+        )
+
+    if A.dtype == torch.float8_e4m3fn:
+        kernel = _w8a8_block_fp8_matmul
+    else:
+        raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.")
+
+    kernel[grid](
+        A,
+        B,
+        C,
+        As,
+        Bs,
+        M,
+        N,
+        K,
+        block_n,
+        block_k,
+        A.stride(-2),
+        A.stride(-1),
+        B.stride(1),
+        B.stride(0),
+        C.stride(-2),
+        C.stride(-1),
+        As.stride(-2),
+        As.stride(-1),
+        Bs.stride(1),
+        Bs.stride(0),
+        **config,
+    )
+
+    return C
+
+
+def get_configs_compute_bound():
+    configs = []
+    for num_stages in [2, 3, 4, 5]:
+        for block_m in [16, 32, 64, 128, 256]:
+            for block_k in [64, 128]:
+                for block_n in [32, 64, 128, 256]:
+                    for num_warps in [4, 8]:
+                        for group_size in [1, 16, 32, 64]:
+                            configs.append(
+                                {
+                                    "BLOCK_SIZE_M": block_m,
+                                    "BLOCK_SIZE_N": block_n,
+                                    "BLOCK_SIZE_K": block_k,
+                                    "GROUP_SIZE_M": group_size,
+                                    "num_warps": num_warps,
+                                    "num_stages": num_stages,
+                                }
+                            )
+    return configs
+
+
+def get_weight_shapes(tp_size):
+    # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3.
+    # Modify them, if you tune for another different model.
+    # cannot TP
+    total = [
+        (512 + 64, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (7168, 16384),
+        (7168, 18432),
+    ]
+    # N can TP
+    n_tp = [
+        (18432 * 2, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (24576, 1536),
+        (12288, 7168),
+        (4096, 7168),
+    ]
+    # K can TP
+    k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)]
+
+    weight_shapes = []
+    for t in total:
+        weight_shapes.append(t)
+    for n_t in n_tp:
+        new_t = (n_t[0] // tp_size, n_t[1])
+        weight_shapes.append(new_t)
+    for k_t in k_tp:
+        new_t = (k_t[0], k_t[1] // tp_size)
+        weight_shapes.append(new_t)
+    return weight_shapes
+
+
+def benchmark_config(
+    A, B, As, Bs, block_size, config, out_dtype=torch.float16, num_iters=10
+):
+    def run():
+        w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
+
+    torch.cuda.synchronize()
+    # JIT complication & warmup
+    for _ in range(5):
+        run()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        run()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    return avg
+
+
+def tune(M, N, K, block_size, out_dtype, search_space, input_type):
+    factor_for_scale = 1e-2
+
+    if input_type == "fp8":
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        A_fp32 = (
+            (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max
+        )
+        A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        B_fp32 = (
+            (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 * fp8_max
+        )
+        B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    else:
+        raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.")
+
+    block_n, block_k = block_size[0], block_size[1]
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+
+    As = torch.rand(M, k_tiles, dtype=torch.float32, device="cuda") * factor_for_scale
+    Bs = (
+        torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda")
+        * factor_for_scale
+    )
+
+    best_config = None
+    best_time = float("inf")
+    for config in tqdm(search_space):
+        try:
+            kernel_time = benchmark_config(
+                A,
+                B,
+                As,
+                Bs,
+                block_size,
+                config,
+                out_dtype,
+                num_iters=10,
+            )
+        except triton.runtime.autotuner.OutOfResources:
+            # Some configurations may be invalid and fail to compile.
+            continue
+
+        if kernel_time < best_time:
+            best_time = kernel_time
+            best_config = config
+    now = datetime.now()
+    print(f"{now.ctime()}] Completed tuning for batch_size={M}")
+    assert best_config is not None
+    return best_config
+
+
+def save_configs(
+    N,
+    K,
+    block_n,
+    block_k,
+    configs,
+    save_path,
+    input_type="fp8",
+) -> None:
+    os.makedirs(save_path, exist_ok=True)
+    device_name = current_platform.get_device_name().replace(" ", "_")
+    json_file_name = (
+        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
+        f"block_shape=[{block_n},{block_k}].json"
+    )
+
+    config_file_path = os.path.join(save_path, json_file_name)
+    print(f"Writing best config to {config_file_path}...")
+
+    with open(config_file_path, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def tune_on_gpu(args_dict):
+    """Run tuning on a specific GPU."""
+    gpu_id = args_dict["gpu_id"]
+    batch_sizes = args_dict["batch_sizes"]
+    weight_shapes = args_dict["weight_shapes"]
+    args = args_dict["args"]
+
+    torch.cuda.set_device(gpu_id)
+    print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
+
+    block_n = args.block_n
+    block_k = args.block_k
+    out_dtype = DTYPE_MAP[args.out_dtype]
+    save_path = args.save_path
+    input_type = args.input_type
+
+    search_space = get_configs_compute_bound()
+    search_space = [
+        config for config in search_space if block_k % config["BLOCK_SIZE_K"] == 0
+    ]
+
+    start = time.time()
+    for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"):
+        N, K = shape[0], shape[1]
+        print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`")
+        benchmark_results = [
+            tune(
+                batch_size,
+                N,
+                K,
+                [block_n, block_k],
+                out_dtype,
+                search_space,
+                input_type,
+            )
+            for batch_size in tqdm(batch_sizes, desc=f"GPU {gpu_id} - Batch sizes")
+        ]
+        best_configs = {M: config for M, config in zip(batch_sizes, benchmark_results)}
+        save_configs(N, K, block_n, block_k, best_configs, save_path, input_type)
+
+    end = time.time()
+    print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds")
+
+
+def distribute_batch_sizes(batch_sizes, num_gpus):
+    """Distribute batch sizes across available GPUs."""
+    batches_per_gpu = []
+    for i in range(num_gpus):
+        start_idx = i * len(batch_sizes) // num_gpus
+        end_idx = (i + 1) * len(batch_sizes) // num_gpus
+        batches_per_gpu.append(batch_sizes[start_idx:end_idx])
+    return batches_per_gpu
+
+
+def main(args):
+    print(args)
+    num_gpus = torch.cuda.device_count()
+    if num_gpus == 0:
+        raise RuntimeError("No GPU available for tuning")
+    print(f"Found {num_gpus} GPUs for parallel tuning")
+
+    torch.cuda.init()
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+        num_gpus = 1  # If only one batch size, use only one GPU
+
+    weight_shapes = get_weight_shapes(args.tp_size)
+
+    batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus)
+
+    process_args = []
+    for gpu_id in range(num_gpus):
+        process_args.append(
+            {
+                "gpu_id": gpu_id,
+                "batch_sizes": batches_per_gpu[gpu_id],
+                "weight_shapes": weight_shapes,  # Each GPU processes all weight shapes
+                "args": args,
+            }
+        )
+
+    ctx = mp.get_context("spawn")
+    with ctx.Pool(num_gpus) as pool:
+        pool.map(tune_on_gpu, process_args)
+
+    print("Multi-GPU tuning completed")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="""
+Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1:
+    python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8
+Then copy to model_executor/layers/quantization/utils/configs
+        """,
+        formatter_class=argparse.RawTextHelpFormatter,
+    )
+
+    parser.add_argument("--tp-size", "-tp", type=int, default=8)
+    parser.add_argument("--input-type", type=str, choices=["fp8"], default="fp8")
+    parser.add_argument(
+        "--out-dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16", "half"],
+        default="float16",
+    )
+    parser.add_argument("--block-n", type=int, default=128)
+    parser.add_argument("--block-k", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--save-path", type=str, default="./")
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/deepgemm/README.md

+# DeepSeek DeepGEMM Kernels Benchmark
+
+This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels.
+
+Currently this just includes dense GEMMs and only works on Hopper GPUs.
+
+## Setup
+
+You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory:
+
+```
+git clone --recursive https://github.com/deepseek-ai/DeepGEMM
+cd DeepGEMM
+python setup.py install
+uv pip install -e .
+```
+
+## Usage
+
+```
+python benchmark_fp8_block_dense_gemm.py
+INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda.
+===== STARTING FP8 GEMM BENCHMARK =====
+PyTorch version: 2.5.1+cu124
+CUDA version: 12.4
+Triton version: 3.1.0
+Using device: NVIDIA H100 80GB HBM3
+WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+
+===== PERFORMANCE COMPARISON =====
+
+DeepGEMM Implementation:
++------+-------+-------+-----------+--------+--------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   |
++------+-------+-------+-----------+--------+--------+
+|    8 |  4096 |  7168 | 102.9     | 4.6    | 286.4  |
+|    8 |  7168 | 18432 | 70.8      | 29.8   | 1868.8 |
+|    8 | 18432 |  7168 | 69.3      | 30.5   | 1911.8 |
+|   64 |  4096 |  7168 | 69.1      | 54.4   | 439.0  |
+|   64 |  7168 | 18432 | 69.4      | 243.6  | 1933.6 |
+|   64 | 18432 |  7168 | 70.4      | 240.3  | 1917.2 |
+|   64 | 24576 |  1536 | 70.1      | 68.9   | 584.6  |
+|   64 | 32768 |   512 | 68.4      | 31.4   | 307.1  |
+|   64 |  7168 | 16384 | 69.5      | 216.3  | 1718.5 |
+|  128 |  4096 |  7168 | 141.1     | 53.3   | 222.1  |
+|  128 |  7168 | 18432 | 71.9      | 470.5  | 1896.1 |
+|  128 | 18432 |  7168 | 69.3      | 488.2  | 1988.2 |
+| 1024 |  4096 |  7168 | 89.7      | 670.1  | 502.5  |
+| 1024 | 18432 |  7168 | 279.0     | 969.8  | 635.2  |
+| 2048 |  4096 |  7168 | 175.1     | 687.0  | 347.4  |
+| 4096 |  4096 |  7168 | 335.4     | 717.0  | 275.1  |
++------+-------+-------+-----------+--------+--------+
+
+vLLM Triton Implementation:
++------+-------+-------+-----------+--------+--------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  |
++------+-------+-------+-----------+--------+--------+--------------+
+|    8 |  4096 |  7168 | 74.0      | 6.3    | 398.2  | 1.39x faster |
+|    8 |  7168 | 18432 | 89.6      | 23.6   | 1478.1 | 0.79x slower |
+|    8 | 18432 |  7168 | 113.2     | 18.7   | 1170.4 | 0.61x slower |
+|   64 |  4096 |  7168 | 79.4      | 47.3   | 382.2  | 0.87x slower |
+|   64 |  7168 | 18432 | 98.5      | 171.7  | 1363.0 | 0.70x slower |
+|   64 | 18432 |  7168 | 119.5     | 141.5  | 1129.4 | 0.59x slower |
+|   64 | 24576 |  1536 | 37.6      | 128.4  | 1089.7 | 1.86x faster |
+|   64 | 32768 |   512 | 38.7      | 55.5   | 542.6  | 1.77x faster |
+|   64 |  7168 | 16384 | 86.1      | 174.5  | 1386.4 | 0.81x slower |
+|  128 |  4096 |  7168 | 90.7      | 82.9   | 345.4  | 1.56x faster |
+|  128 |  7168 | 18432 | 144.0     | 234.9  | 946.9  | 0.50x slower |
+|  128 | 18432 |  7168 | 229.5     | 147.4  | 600.1  | 0.30x slower |
+| 1024 |  4096 |  7168 | 242.3     | 248.2  | 186.1  | 0.37x slower |
+| 1024 | 18432 |  7168 | 897.8     | 301.4  | 197.4  | 0.31x slower |
+| 2048 |  4096 |  7168 | 463.0     | 259.7  | 131.4  | 0.38x slower |
+| 4096 |  4096 |  7168 | 901.8     | 266.7  | 102.3  | 0.37x slower |
++------+-------+-------+-----------+--------+--------+--------------+
+
+vLLM CUTLASS Implementation:
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  | vs Triton    |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+|    8 |  4096 |  7168 | 34.6      | 13.6   | 852.3  | 2.98x faster | 2.14x faster |
+|    8 |  7168 | 18432 | 78.9      | 26.8   | 1677.3 | 0.90x slower | 1.13x faster |
+|    8 | 18432 |  7168 | 81.2      | 26.0   | 1631.1 | 0.85x slower | 1.39x faster |
+|   64 |  4096 |  7168 | 36.9      | 101.9  | 822.9  | 1.87x faster | 2.15x faster |
+|   64 |  7168 | 18432 | 87.4      | 193.4  | 1535.2 | 0.79x slower | 1.13x faster |
+|   64 | 18432 |  7168 | 85.0      | 199.0  | 1587.6 | 0.83x slower | 1.41x faster |
+|   64 | 24576 |  1536 | 28.0      | 172.8  | 1465.8 | 2.51x faster | 1.35x faster |
+|   64 | 32768 |   512 | 28.8      | 74.5   | 728.5  | 2.37x faster | 1.34x faster |
+|   64 |  7168 | 16384 | 77.9      | 193.0  | 1532.8 | 0.89x slower | 1.11x faster |
+|  128 |  4096 |  7168 | 39.1      | 192.4  | 802.0  | 3.61x faster | 2.32x faster |
+|  128 |  7168 | 18432 | 93.7      | 360.8  | 1454.2 | 0.77x slower | 1.54x faster |
+|  128 | 18432 |  7168 | 85.7      | 394.8  | 1608.0 | 0.81x slower | 2.68x faster |
+| 1024 |  4096 |  7168 | 99.7      | 603.1  | 452.2  | 0.90x slower | 2.43x faster |
+| 1024 | 18432 |  7168 | 331.3     | 816.7  | 534.9  | 0.84x slower | 2.71x faster |
+| 2048 |  4096 |  7168 | 198.3     | 606.6  | 306.7  | 0.88x slower | 2.34x faster |
+| 4096 |  4096 |  7168 | 392.2     | 613.2  | 235.3  | 0.86x slower | 2.30x faster |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+
+===== AVERAGE PERFORMANCE =====
++----------------+------------+----------+---------------+
+| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) |
++----------------+------------+----------+---------------+
+| DeepGEMM       | 310.98     | 1052.10  | 0.11          |
+| vLLM Triton    | 144.30     | 715.60   | 0.23          |
+| vLLM CUTLASS   | 286.78     | 1076.67  | 0.11          |
++----------------+------------+----------+---------------+
+
+===== AVERAGE SPEEDUPS =====
++-----------------------------+--------------+
+| Comparison                  | Speedup      |
++-----------------------------+--------------+
+| DeepGEMM vs vLLM Triton     | 1.71x faster |
+| DeepGEMM vs vLLM CUTLASS    | 0.94x slower |
+| vLLM CUTLASS vs vLLM Triton | 1.84x faster |
++-----------------------------+--------------+
+
+===== ACCURACY COMPARISON =====
++----------------+-----------------------+
+| Implementation | Avg Diff vs Reference |
++----------------+-----------------------+
+| DeepGEMM       | 0.000684              |
+| vLLM Triton    | 0.000684              |
+| vLLM CUTLASS   | 0.000684              |
++----------------+-----------------------+
+```

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# fmt: off
+# ruff: noqa: E501
+import time
+
+# Import DeepGEMM functions
+import deep_gemm
+import torch
+from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor
+
+# Import vLLM functions
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    per_token_group_quant_fp8,
+    w8a8_block_fp8_matmul,
+)
+from vllm.triton_utils import triton
+
+
+# Copied from
+# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9
+def per_token_cast_to_fp8(
+        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert tensor to FP8 format with per-token scaling."""
+    assert x.dim() == 2 and x.size(1) % 128 == 0
+    m, n = x.shape
+    x_view = x.view(m, -1, 128)
+    x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4)
+    return (x_view * (448.0 / x_amax.unsqueeze(2))).to(
+        torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1)
+
+
+# Copied from
+# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17
+def per_block_cast_to_fp8(
+        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert tensor to FP8 format with per-block scaling."""
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128),
+                           dtype=x.dtype,
+                           device=x.device)
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
+    return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (
+        x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+
+
+def benchmark_shape(m: int,
+                    n: int,
+                    k: int,
+                    warmup: int = 100,
+                    repeat: int = 10000,
+                    verbose: bool = False) -> dict:
+    """Benchmark all implementations for a specific (m, n, k) shape."""
+    if verbose:
+        print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===")
+
+    # Create test tensors
+    A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
+    B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16)
+
+    # Reference result in BF16
+    torch.cuda.synchronize()
+    C_ref = A @ B.t()
+
+    # Pre-quantize B for all implementations
+    # (weights can be pre-quantized offline)
+    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B)
+    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B)
+
+    # Block size configuration
+    block_size = [128, 128]
+
+    # Pre-quantize A for all implementations
+    A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
+    A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
+    C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
+    A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
+    A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
+        A, block_size[1], column_major_scales=True)
+
+    # === DeepGEMM Implementation ===
+    def deepgemm_gemm():
+        deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm),
+                                       (B_deepgemm, B_scale_deepgemm),
+                                       C_deepgemm)
+        return C_deepgemm
+
+    # === vLLM Triton Implementation ===
+    def vllm_triton_gemm():
+        return w8a8_block_fp8_matmul(A_vllm,
+                                     B_vllm,
+                                     A_scale_vllm,
+                                     B_scale_vllm,
+                                     block_size,
+                                     output_dtype=torch.bfloat16)
+
+    # === vLLM CUTLASS Implementation ===
+    def vllm_cutlass_gemm():
+        return ops.cutlass_scaled_mm(A_vllm_cutlass,
+                                     B_vllm.T,
+                                     scale_a=A_scale_vllm_cutlass,
+                                     scale_b=B_scale_vllm.T,
+                                     out_dtype=torch.bfloat16)
+
+    # Run correctness check first
+    if verbose:
+        print("Running correctness check...")
+    C_deepgemm = deepgemm_gemm()
+    C_vllm_triton = vllm_triton_gemm()
+    C_vllm_cutlass = vllm_cutlass_gemm()
+
+    deepgemm_diff = calc_diff(C_deepgemm, C_ref)
+    vllm_triton_diff = calc_diff(C_vllm_triton, C_ref)
+    vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref)
+
+    if verbose:
+        print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}")
+        print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}")
+        print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}")
+        print("vLLM Triton vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}")
+        print("vLLM CUTLASS vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}")
+
+    # Benchmark implementations
+    implementations = {
+        "DeepGEMM": deepgemm_gemm,
+        "vLLM Triton": vllm_triton_gemm,
+        "vLLM CUTLASS": vllm_cutlass_gemm
+    }
+
+    benchmark_results = {
+        "shape": {
+            "m": m,
+            "n": n,
+            "k": k
+        },
+        "implementations": {}
+    }
+
+    for name, func in implementations.items():
+        # Warmup
+        for _ in range(warmup):
+            func()
+            torch.cuda.synchronize()
+
+        # Timing loop
+        torch.cuda.synchronize()
+        start = time.time()
+        for _ in range(repeat):
+            func()
+        torch.cuda.synchronize()
+        end = time.time()
+
+        # Calculate timing and TFLOPS
+        avg_time_ms = (end - start) / repeat * 1000
+        avg_time_us = avg_time_ms * 1000
+        tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12
+        gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3)
+
+        benchmark_results["implementations"][name] = {
+            "time_ms": avg_time_ms,
+            "time_us": avg_time_us,
+            "tflops": tflops,
+            "gb_s": gb_s,
+            "diff": {
+                "DeepGEMM":
+                0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm),
+                "Reference":
+                deepgemm_diff if name == "DeepGEMM" else
+                (vllm_triton_diff
+                 if name == "vLLM Triton" else vllm_cutlass_diff)
+            }
+        }
+
+        if verbose:
+            print(
+                f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s"
+            )
+
+    # Calculate speedups
+    baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"]
+    for name, data in benchmark_results["implementations"].items():
+        if name != "DeepGEMM":
+            speedup = baseline / data["time_ms"]
+            benchmark_results["implementations"][name][
+                "speedup_vs_deepgemm"] = speedup
+            if verbose:
+                print(f"DeepGEMM is {1/speedup:.2f}x "
+                      f"{'faster' if 1/speedup > 1 else 'slower'} than {name}")
+
+    vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][
+        "time_ms"]
+    vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][
+        "time_ms"]
+    cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time
+    benchmark_results["implementations"]["vLLM CUTLASS"][
+        "speedup_vs_triton"] = cutlass_vs_triton
+    if verbose:
+        print(
+            f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x "
+            f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton"
+        )
+
+    return benchmark_results
+
+
+def format_table_row(values, widths):
+    """Format a row with specified column widths."""
+    return "| " + " | ".join(f"{val:{w}}"
+                             for val, w in zip(values, widths)) + " |"
+
+
+def print_table(headers, rows, title=None):
+    """Print a table with headers and rows."""
+    if title:
+        print(f"\n{title}")
+
+    # Calculate column widths based on headers and data
+    widths = [
+        max(len(str(h)), max(len(str(row[i])) for row in rows))
+        for i, h in enumerate(headers)
+    ]
+
+    # Create separator line
+    separator = "+-" + "-+-".join("-" * w for w in widths) + "-+"
+
+    # Print table
+    print(separator)
+    print(format_table_row(headers, widths))
+    print(separator)
+    for row in rows:
+        print(format_table_row(row, widths))
+    print(separator)
+
+
+def format_speedup(value):
+    """Format speedup value with indicator if it's faster or slower."""
+    return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}"
+
+
+def run_benchmarks(verbose: bool = False):
+    """Run benchmarks for a set of common shapes."""
+    print("===== STARTING FP8 GEMM BENCHMARK =====")
+
+    # Make sure we're using the GPU
+    if not torch.cuda.is_available():
+        print("CUDA not available! Tests require GPU.")
+        return
+
+    # Print system information
+    print(f"PyTorch version: {torch.__version__}")
+    print(f"CUDA version: {torch.version.cuda}")
+    print(f"Triton version: {triton.__version__}")
+    print(f"Using device: {torch.cuda.get_device_name()}")
+
+    # Enable TF32 for better performance
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+    # Set seeds for reproducibility
+    torch.manual_seed(42)
+    torch.cuda.manual_seed(42)
+
+    # Define benchmark shapes (m, n, k)
+    shapes = [
+        (8, 4096, 7168),
+        (8, 7168, 18432),
+        (8, 18432, 7168),
+        (64, 4096, 7168),
+        (64, 7168, 18432),
+        (64, 18432, 7168),
+        (64, 24576, 1536),
+        (64, 32768, 512),
+        (64, 7168, 16384),
+        (128, 4096, 7168),
+        (128, 7168, 18432),
+        (128, 18432, 7168),
+        (1024, 4096, 7168),
+        (1024, 18432, 7168),
+        (2048, 4096, 7168),
+        (4096, 4096, 7168),
+    ]
+    shapes = [
+        # (64, 2112, 7168),
+        (64, 24576, 1536),
+        (64, 32768, 512),
+        (64, 7168, 16384),
+        (64, 4096, 7168),
+        (64, 7168, 2048),
+        # (128, 2112, 7168),
+        (128, 24576, 1536),
+        (128, 32768, 512),
+        (128, 7168, 16384),
+        (128, 4096, 7168),
+        (128, 7168, 2048),
+        # (4096, 2112, 7168),
+        (4096, 24576, 1536),
+        (4096, 32768, 512),
+        (4096, 7168, 16384),
+        (4096, 4096, 7168),
+        (4096, 7168, 2048),
+    ]
+
+    all_results = []
+    for m, n, k in shapes:
+        result = benchmark_shape(m, n, k, verbose=verbose)
+        all_results.append(result)
+
+    # Print results in a nicely formatted table
+    print("\n===== PERFORMANCE COMPARISON =====")
+
+    # Print DeepGEMM table
+    deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"]
+    deepgemm_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["DeepGEMM"]
+        deepgemm_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}"
+        ])
+
+    print_table(deepgemm_headers,
+                deepgemm_rows,
+                title="DeepGEMM Implementation:")
+
+    # Print vLLM Triton table
+    triton_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM"
+    ]
+    triton_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["vLLM Triton"]
+        speedup = impl_data.get("speedup_vs_deepgemm", 1.0)
+        triton_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(speedup)
+        ])
+
+    print_table(triton_headers,
+                triton_rows,
+                title="vLLM Triton Implementation:")
+
+    # Print vLLM CUTLASS table
+    cutlass_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM",
+        "vs Triton"
+    ]
+    cutlass_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["vLLM CUTLASS"]
+        vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0)
+        vs_triton = impl_data.get("speedup_vs_triton", 1.0)
+        cutlass_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(vs_deepgemm),
+            format_speedup(vs_triton)
+        ])
+
+    print_table(cutlass_headers,
+                cutlass_rows,
+                title="vLLM CUTLASS Implementation:")
+
+    # Calculate and print averages
+    print("\n===== AVERAGE PERFORMANCE =====")
+
+    implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"]
+    avg_metrics = {
+        impl: {
+            "tflops": 0,
+            "gb_s": 0,
+            "time_ms": 0
+        }
+        for impl in implementations
+    }
+
+    for result in all_results:
+        for impl in implementations:
+            impl_data = result["implementations"][impl]
+            avg_metrics[impl]["tflops"] += impl_data["tflops"]
+            avg_metrics[impl]["gb_s"] += impl_data["gb_s"]
+            avg_metrics[impl]["time_ms"] += impl_data["time_ms"]
+
+    num_shapes = len(all_results)
+    avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"]
+    avg_rows = []
+
+    for impl in implementations:
+        avg_tflops = avg_metrics[impl]["tflops"] / num_shapes
+        avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes
+        avg_time = avg_metrics[impl]["time_ms"] / num_shapes
+        avg_rows.append([
+            impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}"
+        ])
+
+    print_table(avg_headers, avg_rows)
+
+    # Calculate average speedups
+    avg_speedups = {
+        "DeepGEMM vs vLLM Triton": 0,
+        "DeepGEMM vs vLLM CUTLASS": 0,
+        "vLLM CUTLASS vs vLLM Triton": 0
+    }
+
+    for result in all_results:
+        deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"]
+        vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"]
+        vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][
+            "time_ms"]
+
+        avg_speedups[
+            "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time
+        avg_speedups[
+            "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time
+        avg_speedups[
+            "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time
+
+    print("\n===== AVERAGE SPEEDUPS =====")
+    speedup_headers = ["Comparison", "Speedup"]
+    speedup_rows = []
+    for comparison, total in avg_speedups.items():
+        avg_speedup = total / num_shapes
+        status = "faster" if avg_speedup > 1 else "slower"
+        speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"])
+
+    print_table(speedup_headers, speedup_rows)
+
+    # Average accuracy comparison
+    print("\n===== ACCURACY COMPARISON =====")
+    avg_diff = {impl: 0 for impl in implementations}
+
+    for result in all_results:
+        for impl in implementations:
+            avg_diff[impl] += result["implementations"][impl]["diff"][
+                "Reference"]
+
+    diff_headers = ["Implementation", "Avg Diff vs Reference"]
+    diff_rows = []
+    for impl in implementations:
+        diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"])
+
+    print_table(diff_headers, diff_rows)
+
+
+if __name__ == "__main__":
+    run_benchmarks(verbose=False)

benchmarks/kernels/graph_machete_bench.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 import math
 import pickle
-import re
 from collections import defaultdict
-from typing import List
 
 import matplotlib.pyplot as plt
 import pandas as pd
+import regex as re
 import seaborn as sns
 from torch.utils.benchmark import Measurement as TMeasurement
 
@@ -13,17 +15,19 @@ from vllm.utils import FlexibleArgumentParser
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description='Benchmark the latency of processing a single batch of '
-        'requests till completion.')
-    parser.add_argument('filename', type=str)
+        description="Benchmark the latency of processing a single batch of "
+        "requests till completion."
+    )
+    parser.add_argument("filename", type=str)
 
     args = parser.parse_args()
 
-    with open(args.filename, 'rb') as f:
-        data: List[TMeasurement] = pickle.load(f)
+    with open(args.filename, "rb") as f:
+        data = pickle.load(f)
+        raw_results: list[TMeasurement] = data["results"]
 
     results = defaultdict(lambda: list())
-    for v in data:
+    for v in raw_results:
         result = re.search(r"MKN=\(\d+x(\d+x\d+)\)", v.task_spec.sub_label)
         if result is not None:
             KN = result.group(1)
@@ -36,11 +40,7 @@ if __name__ == "__main__":
             raise Exception("MKN not found")
 
         kernel = v.task_spec.description
-        results[KN].append({
-            "kernel": kernel,
-            "batch_size": M,
-            "median": v.median
-        })
+        results[KN].append({"kernel": kernel, "batch_size": M, "median": v.median})
 
     rows = int(math.ceil(len(results) / 2))
     fig, axs = plt.subplots(rows, 2, figsize=(12, 5 * rows))
@@ -48,14 +48,16 @@ if __name__ == "__main__":
     for axs_idx, (shape, data) in enumerate(results.items()):
         plt.sca(axs[axs_idx])
         df = pd.DataFrame(data)
-        sns.lineplot(data=df,
-                     x="batch_size",
-                     y="median",
-                     hue="kernel",
-                     style="kernel",
-                     markers=True,
-                     dashes=False,
-                     palette="Dark2")
+        sns.lineplot(
+            data=df,
+            x="batch_size",
+            y="median",
+            hue="kernel",
+            style="kernel",
+            markers=True,
+            dashes=False,
+            palette="Dark2",
+        )
         plt.title(f"Shape: {shape}")
         plt.ylabel("time (median, s)")
     plt.tight_layout()

benchmarks/kernels/utils.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import dataclasses
+from collections.abc import Iterable
+from typing import Any, Callable, Optional
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+
+
+@dataclasses.dataclass
+class CudaGraphBenchParams:
+    num_ops_in_cuda_graph: int
+
+
+@dataclasses.dataclass
+class ArgPool:
+    """
+    When some argument of the benchmarking function is annotated with this type,
+    the benchmarking class (BenchMM) will collapse the argument to a pick a
+    single value from the given list of values, during function invocation.
+    For every invocation during a benchmarking run, it will choose a
+    different value from the list.
+    """
+
+    values: Iterable[Any]
+
+    def __getitem__(self, index):
+        return self.values[index]
+
+
+class Bench:
+    class ArgsIterator:
+        def __init__(self, args_list, kwargs_list):
+            assert len(args_list) == len(kwargs_list)
+            self.args_list = args_list
+            self.kwargs_list = kwargs_list
+            self.n = len(self.args_list)
+            self.idx = 0
+
+        def __next__(self):
+            while True:
+                yield (self.args_list[self.idx], self.kwargs_list[self.idx])
+                self.idx += 1
+                self.idx = self.idx % self.n
+
+        def reset(self):
+            self.idx = 0
+
+        @property
+        def n_args(self):
+            return self.n
+
+    def __init__(
+        self,
+        cuda_graph_params: Optional[CudaGraphBenchParams],
+        label: str,
+        sub_label: str,
+        description: str,
+        fn: Callable,
+        *args,
+        **kwargs,
+    ):
+        self.cuda_graph_params = cuda_graph_params
+        self.use_cuda_graph = self.cuda_graph_params is not None
+        self.label = label
+        self.sub_label = sub_label
+        self.description = description
+        self.fn = fn
+
+        # Process args
+        self._args = args
+        self._kwargs = kwargs
+        self.args_list, self.kwargs_list = self.collapse_argpool(*args, **kwargs)
+        self.args_iterator = self.ArgsIterator(self.args_list, self.kwargs_list)
+
+        # Cudagraph runner
+        self.g = None
+        if self.use_cuda_graph:
+            self.g = self.get_cuda_graph_runner()
+
+        # benchmark run params
+        self.min_run_time = 1
+
+    def collapse_argpool(self, *args, **kwargs):
+        argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [
+            arg for arg in kwargs.values() if isinstance(arg, ArgPool)
+        ]
+        if len(argpool_args) == 0:
+            return [args], [kwargs]
+
+        # Make sure all argpools are of the same size
+        argpool_size = len(argpool_args[0].values)
+        assert all([argpool_size == len(arg.values) for arg in argpool_args])
+
+        # create copies of the args
+        args_list = []
+        kwargs_list = []
+        for _ in range(argpool_size):
+            args_list.append(args)
+            kwargs_list.append(kwargs.copy())
+
+        for i in range(argpool_size):
+            # collapse args; Just pick the ith value
+            args_list[i] = tuple(
+                [arg[i] if isinstance(arg, ArgPool) else arg for arg in args_list[i]]
+            )
+
+            # collapse kwargs
+            kwargs_i = kwargs_list[i]
+            arg_pool_keys = [k for k, v in kwargs_i.items() if isinstance(v, ArgPool)]
+            for k in arg_pool_keys:
+                # again just pick the ith value
+                kwargs_i[k] = kwargs_i[k][i]
+            kwargs_list[i] = kwargs_i
+
+        return args_list, kwargs_list
+
+    def get_cuda_graph_runner(self):
+        assert self.use_cuda_graph
+        assert self.args_iterator is not None
+
+        num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph
+
+        # warmup
+        args_it = self.args_iterator.__next__()
+        for _ in range(2):
+            args, kwargs = next(args_it)
+            self.fn(*args, **kwargs)
+
+        self.args_iterator.reset()
+        args_it = self.args_iterator.__next__()
+        stream = torch.cuda.Stream()
+        with torch.cuda.stream(stream):
+            g = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(g):
+                for _ in range(num_graph_ops):
+                    args, kwargs = next(args_it)
+                    self.fn(*args, **kwargs)
+        return g
+
+    def run_cudagrah(self) -> TMeasurement:
+        assert self.use_cuda_graph
+        globals = {"g": self.g}
+
+        return TBenchmark.Timer(
+            stmt="g.replay()",
+            globals=globals,
+            label=(
+                f"{self.label}"
+                f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops"
+            ),
+            sub_label=self.sub_label,
+            description=self.description,
+        ).blocked_autorange(min_run_time=self.min_run_time)
+
+    def run_eager(self) -> TMeasurement:
+        setup = None
+        stmt = None
+        globals = None
+
+        has_arg_pool = self.args_iterator.n_args > 1
+        if has_arg_pool:
+            setup = """
+                    args_iterator.reset()
+                    args_it = args_iterator.__next__()
+                    """
+            stmt = """
+                    args, kwargs = next(args_it)
+                    fn(*args, **kwargs)
+                    """
+            globals = {"fn": self.fn, "args_iterator": self.args_iterator}
+        else:
+            # no arg pool. Just use the args and kwargs directly
+            self.args_iterator.reset()
+            args_it = self.args_iterator.__next__()
+            args, kwargs = next(args_it)
+
+            setup = ""
+            stmt = """
+                    fn(*args, **kwargs)
+                   """
+            globals = {"fn": self.fn, "args": args, "kwargs": kwargs}
+
+        return TBenchmark.Timer(
+            stmt=stmt,
+            setup=setup,
+            globals=globals,
+            label=self.label,
+            sub_label=self.sub_label,
+            description=self.description,
+        ).blocked_autorange(min_run_time=self.min_run_time)
+
+    def run(self) -> TMeasurement:
+        timer = None
+        if self.use_cuda_graph:  # noqa SIM108
+            timer = self.run_cudagrah()
+        else:
+            timer = self.run_eager()
+        if not timer.meets_confidence() or timer.has_warnings:
+            print("Doesn't meet confidence - re-running bench ...")
+            return self.run()
+        return timer
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        if exc_type:
+            print(f"exc type {exc_type}")
+            print(f"exc value {exc_value}")
+            print(f"exc traceback {traceback}")

benchmarks/kernels/weight_shapes.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:
@@ -40,4 +43,56 @@ WEIGHT_SHAPES = {
         ([8192, 57344], 1),
         ([28672, 8192], 0),
     ],
+    "meta-llama/Llama-3.1-405b-hf": [
+        ([16384, 18432], 1),
+        ([16384, 16384], 0),
+        ([16384, 106496], 1),
+        ([53248, 16384], 0),
+    ],
+    "meta-llama/Llama-3.1-8B-Instruct": [
+        ([4096, 6144], 1),
+        ([4096, 4096], 0),
+        ([4096, 28672], 1),
+        ([14336, 4096], 0),
+    ],
+    "meta-llama/Llama-3.3-70B-Instruct": [
+        ([8192, 10240], 1),
+        ([8192, 8192], 0),
+        ([8192, 57344], 1),
+        ([28672, 8192], 0),
+    ],
+    "mistralai/Mistral-Large-Instruct-2407": [
+        ([12288, 14336], 1),
+        ([12288, 12288], 0),
+        ([12288, 57344], 1),
+        ([28672, 12288], 0),
+    ],
+    "Qwen/Qwen2.5-7B-Instruct": [
+        ([3584, 4608], 1),
+        ([3584, 3584], 0),
+        ([3584, 37888], 1),
+        ([18944, 3584], 0),
+    ],
+    "Qwen/Qwen2.5-32B-Instruct": [
+        ([5120, 7168], 1),
+        ([5120, 5120], 0),
+        ([5120, 55296], 1),
+        ([27648, 5120], 0),
+    ],
+    "Qwen/Qwen2.5-72B-Instruct": [
+        ([8192, 10240], 1),
+        ([8192, 8192], 0),
+        ([8192, 59136], 1),
+        ([29568, 8192], 0),
+    ],
+    "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [
+        ([2048, 3072], 1),
+        ([2048, 4096], 1),
+        ([2048, 2048], 0),
+        ([2048, 576], 0),
+        ([2048, 21888], 1),
+        ([10944, 2048], 0),
+        ([2048, 2816], 1),
+        ([1408, 2048], 0),
+    ],
 }

benchmarks/launch_tgi_server.sh

-#!/bin/bash
-
-PORT=8000
-MODEL=$1
-TOKENS=$2
-
-docker run -e HF_TOKEN=$HF_TOKEN --gpus all --shm-size 1g -p $PORT:80 \
-           -v $PWD/data:/data \
-           ghcr.io/huggingface/text-generation-inference:2.2.0 \
-           --model-id $MODEL \
-           --sharded false  \
-           --max-input-length 1024 \
-           --max-total-tokens 2048 \
-           --max-best-of 5 \
-           --max-concurrent-requests 5000 \
-           --max-batch-total-tokens $TOKENS

benchmarks/overheads/benchmark_hashing.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 import cProfile
 import pstats
 
@@ -5,9 +8,8 @@ from vllm import LLM, SamplingParams
 from vllm.utils import FlexibleArgumentParser
 
 # A very long prompt, total number of tokens is about 15k.
-LONG_PROMPT = ["You are an expert in large language models, aren't you?"
-               ] * 1000
-LONG_PROMPT = ' '.join(LONG_PROMPT)
+LONG_PROMPT = ["You are an expert in large language models, aren't you?"] * 1000
+LONG_PROMPT = " ".join(LONG_PROMPT)
 
 
 def main(args):
@@ -16,7 +18,6 @@ def main(args):
         enforce_eager=True,
         enable_prefix_caching=True,
         tensor_parallel_size=args.tensor_parallel_size,
-        use_v2_block_manager=args.use_v2_block_manager,
     )
 
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
@@ -29,35 +30,35 @@ def main(args):
 
     print("------start generating------")
     for i in range(3):
-        profiler.runctx('llm.generate(LONG_PROMPT, sampling_params)',
-                        globals(), locals())
+        profiler.runctx(
+            "llm.generate(LONG_PROMPT, sampling_params)", globals(), locals()
+        )
 
     # analyze the runtime of hashing function
     stats = pstats.Stats(profiler)
-    stats.sort_stats('cumulative')
+    stats.sort_stats("cumulative")
     total_time = 0
     total_calls = 0
     for func in stats.stats:
-        if 'hash_of_block' in func[2]:
+        if "hash_of_block" in func[2]:
             total_time = stats.stats[func][3]
             total_calls = stats.stats[func][0]
     percentage = (total_time / stats.total_tt) * 100
-    print(f"Hashing took {total_time:.2f} seconds,"
-          f"{percentage:.2f}% of the total runtime.")
+    print(
+        f"Hashing took {total_time:.2f} seconds,{percentage:.2f}% of the total runtime."
+    )
 
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description='Benchmark the performance of hashing function in'
-        'automatic prefix caching.')
-    parser.add_argument('--model', type=str, default='lmsys/longchat-7b-16k')
-    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
-    parser.add_argument('--output-len', type=int, default=10)
-    parser.add_argument('--enable-prefix-caching',
-                        action='store_true',
-                        help='enable prefix caching')
-    parser.add_argument('--use-v2-block-manager',
-                        action='store_true',
-                        help='Use BlockSpaceMangerV2')
+        description="Benchmark the performance of hashing function in"
+        "automatic prefix caching."
+    )
+    parser.add_argument("--model", type=str, default="lmsys/longchat-7b-16k")
+    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
+    parser.add_argument("--output-len", type=int, default=10)
+    parser.add_argument(
+        "--enable-prefix-caching", action="store_true", help="enable prefix caching"
+    )
     args = parser.parse_args()
     main(args)

benchmarks/pyproject.toml

+# This local pyproject file is part of the migration from yapf to ruff format.
+# It uses the same core rules as the main pyproject.toml file, but with the
+# following differences:
+# - ruff line length is overridden to 88
+# - deprecated typing ignores (UP006, UP035) have been removed
+
+[tool.ruff]
+line-length = 88
+
+[tool.ruff.lint.per-file-ignores]
+"vllm/third_party/**" = ["ALL"]
+"vllm/version.py" = ["F401"]
+"vllm/_version.py" = ["ALL"]
+
+[tool.ruff.lint]
+select = [
+    # pycodestyle
+    "E",
+    # Pyflakes
+    "F",
+    # pyupgrade
+    "UP",
+    # flake8-bugbear
+    "B",
+    # flake8-simplify
+    "SIM",
+    # isort
+    "I",
+    # flake8-logging-format
+    "G",
+]
+ignore = [
+    # star imports
+    "F405", "F403",
+    # lambda expression assignment
+    "E731",
+    # Loop control variable not used within loop body
+    "B007",
+    # f-string format
+    "UP032",
+    # Can remove once 3.10+ is the minimum Python version
+    "UP007",
+]
+
+[tool.ruff.lint.isort]
+known-first-party = ["vllm"]
+
+[tool.ruff.format]
+docstring-code-format = true
\ No newline at end of file

benchmarks/run_structured_output_benchmark.sh

+#!/bin/bash
+
+# default values
+MODEL=${MODEL:-"Qwen/Qwen2.5-7B-Instruct"}
+BACKEND=${BACKEND:-"vllm"}
+DATASET=${DATASET:-"xgrammar_bench"}
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+OUTPUT_DIR=${OUTPUT_DIR:-"$SCRIPT_DIR/structured_output_benchmark_results"}
+PORT=${PORT:-8000}
+STRUCTURED_OUTPUT_RATIO=${STRUCTURED_OUTPUT_RATIO:-1}
+TOTAL_SECONDS=${TOTAL_SECONDS:-90}
+MAX_NEW_TOKENS=${MAX_NEW_TOKENS:-300}
+TOKENIZER_MODE=${TOKENIZER_MODE:-"auto"}
+
+usage() {
+    echo "Usage: $0 [options]"
+    echo "Options:"
+    echo "  --model MODEL                  Model to benchmark (default: $MODEL)"
+    echo "  --backend BACKEND              Backend to use (default: $BACKEND)" 
+    echo "  --dataset DATASET              Dataset to use (default: $DATASET)"
+    echo "  --max-new-tokens N             Maximum number of tokens to generate (default: $MAX_NEW_TOKENS)"
+    echo "  --output-dir DIR               Output directory for results (default: $OUTPUT_DIR)"
+    echo "  --port PORT                    Port to use (default: $PORT)"
+    echo "  --structured-output-ratio N    Ratio of structured outputs (default: $STRUCTURED_OUTPUT_RATIO)"
+    echo "  --tokenizer-mode MODE          Tokenizer mode to use (default: $TOKENIZER_MODE)"
+    echo "  --total-seconds N              Total seconds to run the benchmark (default: $TOTAL_SECONDS)"
+    echo "  -h, --help                     Show this help message and exit"
+    exit 0
+}
+
+# parse command line arguments
+while [[ $# -gt 0 ]]; do
+  case $1 in
+    --model)
+      MODEL="$2"
+      shift 2
+      ;;
+    --backend)
+      BACKEND="$2"
+      shift 2
+      ;;
+    --dataset)
+      DATASET="$2"
+      shift 2
+      ;;
+    --max-new-tokens)
+      MAX_NEW_TOKENS="$2"
+      shift 2
+      ;;
+    --output-dir)
+      OUTPUT_DIR="$2"
+      shift 2
+      ;;
+    --port)
+      PORT="$2"
+      shift 2
+      ;;
+    --structured-output-ratio)
+      STRUCTURED_OUTPUT_RATIO="$2"
+      shift 2
+      ;;
+    --tokenizer-mode)
+      TOKENIZER_MODE="$2"
+      shift 2
+      ;;
+    --total-seconds)
+      TOTAL_SECONDS="$2"
+      shift 2
+      ;;
+    -h|--help)
+      usage
+      ;;
+    *)
+      echo "Unknown argument: $1\n"
+      usage
+      ;;
+  esac
+done
+
+# Create output directory if it doesn't exist
+mkdir -p "$OUTPUT_DIR"
+
+# Define QPS values to test
+QPS_VALUES=(25 20 15 10 5 1)
+
+# Common parameters
+COMMON_PARAMS="--backend $BACKEND \
+               --model $MODEL \
+               --dataset $DATASET \
+               --structured-output-ratio $STRUCTURED_OUTPUT_RATIO \
+               --save-results \
+               --result-dir $OUTPUT_DIR \
+               --output-len $MAX_NEW_TOKENS \
+               --port $PORT \
+               --tokenizer-mode $TOKENIZER_MODE"
+
+echo "Starting structured output benchmark with model: $MODEL"
+echo "Backend: $BACKEND"
+echo "Dataset: $DATASET"
+echo "Results will be saved to: $OUTPUT_DIR"
+echo "----------------------------------------"
+
+# Run benchmarks with different QPS values
+for qps in "${QPS_VALUES[@]}"; do
+  echo "Running benchmark with QPS: $qps"
+
+  # Get git hash and branch for the filename
+  GIT_HASH=$(git rev-parse --short HEAD 2>/dev/null || echo "unknown")
+  GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown")
+
+  # Construct filename for this run
+  FILENAME="${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json"
+
+  NUM_PROMPTS=$(echo "$TOTAL_SECONDS * $qps" | bc)
+  NUM_PROMPTS=${NUM_PROMPTS%.*}  # Remove fractional part
+  echo "Running benchmark with $NUM_PROMPTS prompts"
+
+  # Run the benchmark
+  python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \
+    --request-rate $qps \
+    --result-filename "$FILENAME" \
+    --num-prompts $NUM_PROMPTS
+
+  echo "Completed benchmark with QPS: $qps"
+  echo "----------------------------------------"
+done
+
+echo "All benchmarks completed!"
+echo "Results saved to: $OUTPUT_DIR"

benchmarks/structured_schemas/structured_schema_1.json

+{
+    "type": "object",
+    "properties": {
+      "name": { "type": "string" },
+      "email": { "type": "string" },
+      "street": { "type": "string" },
+      "city": { "type": "string" },
+      "state": { "type": "string" },
+      "zip": { "type": "string" },
+      "phone": { "type": "string" },
+      "website": { "type": "string" },
+      "company": { "type": "string" },
+      "age": { "type": "integer" }
+    },
+    "required": [
+      "name",
+      "email"
+    ]
+}