Merge tag 'v0.9.0' into v0.9.0-ori

benchmarks/kernels/benchmark_paged_attention.py

@@ -9,8 +9,11 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        create_kv_caches_with_random)
+from vllm.utils import (
+    STR_DTYPE_TO_TORCH_DTYPE,
+    FlexibleArgumentParser,
+    create_kv_caches_with_random,
+)
 
 logger = init_logger(__name__)
 
@@ -38,19 +41,15 @@ def main(
     current_platform.seed_everything(seed)
 
     scale = float(1.0 / (head_size**0.5))
-    query = torch.empty(num_seqs,
-                        num_query_heads,
-                        head_size,
-                        dtype=dtype,
-                        device=device)
+    query = torch.empty(
+        num_seqs, num_query_heads, head_size, dtype=dtype, device=device
+    )
     query.uniform_(-scale, scale)
 
     assert num_query_heads % num_kv_heads == 0
     alibi_slopes = None
     if use_alibi:
-        alibi_slopes = torch.randn(num_query_heads,
-                                   dtype=torch.float,
-                                   device=device)
+        alibi_slopes = torch.randn(num_query_heads, dtype=torch.float, device=device)
 
     seq_lens = [seq_len for _ in range(num_seqs)]
     max_seq_len = max(seq_lens)
@@ -61,24 +60,23 @@ def main(
     block_tables_lst: list[list[int]] = []
     for _ in range(num_seqs):
         block_table = [
-            random.randint(0, NUM_BLOCKS - 1)
-            for _ in range(max_num_blocks_per_seq)
+            random.randint(0, NUM_BLOCKS - 1) for _ in range(max_num_blocks_per_seq)
         ]
         block_tables_lst.append(block_table)
 
-    block_tables = torch.tensor(block_tables_lst,
-                                dtype=torch.int,
-                                device=device)
+    block_tables = torch.tensor(block_tables_lst, dtype=torch.int, device=device)
 
     # Create the KV cache.
-    key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
-                                                            block_size,
-                                                            1,
-                                                            num_kv_heads,
-                                                            head_size,
-                                                            kv_cache_dtype,
-                                                            dtype,
-                                                            device=device)
+    key_caches, value_caches = create_kv_caches_with_random(
+        NUM_BLOCKS,
+        block_size,
+        1,
+        num_kv_heads,
+        head_size,
+        kv_cache_dtype,
+        dtype,
+        device=device,
+    )
     key_cache, value_cache = key_caches[0], value_caches[0]
 
     # Prepare for the paged attention kernel.
@@ -86,11 +84,8 @@ def main(
     if version == "v2":
         if current_platform.is_rocm():
             global PARTITION_SIZE
-            if not args.custom_paged_attn:
-                PARTITION_SIZE = 1024
-            else:
-                PARTITION_SIZE = PARTITION_SIZE_ROCM
-        num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
+            PARTITION_SIZE = 1024 if not args.custom_paged_attn else PARTITION_SIZE_ROCM
+        num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE
         tmp_output = torch.empty(
             size=(num_seqs, num_query_heads, num_partitions, head_size),
             dtype=output.dtype,
@@ -110,9 +105,7 @@ def main(
         start_time = time.perf_counter()
 
         # Using default kv_scale
-        k_scale = v_scale = torch.tensor(1.0,
-                                         dtype=torch.float32,
-                                         device=device)
+        k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device)
 
         for _ in range(num_iters):
             if version == "v1":
@@ -195,30 +188,29 @@ def main(
     print(f"Kernel running time: {latency * 1000000:.3f} us")
 
 
-if __name__ == '__main__':
-    logger.warning("This script benchmarks the paged attention kernel. "
-                   "By default this is no longer used in vLLM inference.")
+if __name__ == "__main__":
+    logger.warning(
+        "This script benchmarks the paged attention kernel. "
+        "By default this is no longer used in vLLM inference."
+    )
 
-    parser = FlexibleArgumentParser(
-        description="Benchmark the paged attention kernel.")
-    parser.add_argument("--version",
-                        type=str,
-                        choices=["v1", "v2"],
-                        default="v2")
+    parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.")
+    parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2")
     parser.add_argument("--batch-size", type=int, default=8)
     parser.add_argument("--seq-len", type=int, default=4096)
     parser.add_argument("--num-query-heads", type=int, default=64)
     parser.add_argument("--num-kv-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("--block-size", type=int, choices=[16, 32], default=16)
     parser.add_argument("--use-alibi", action="store_true")
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["half", "bfloat16", "float"],
-                        default="half")
+    parser.add_argument(
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
+    )
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--profile", action="store_true")
     parser.add_argument(
@@ -228,10 +220,11 @@ if __name__ == '__main__':
         default="auto",
         help="Data type for kv cache storage. If 'auto', will use model "
         "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
-        "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)")
-    parser.add_argument("--custom-paged-attn",
-                        action="store_true",
-                        help="Use custom paged attention")
+        "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)",
+    )
+    parser.add_argument(
+        "--custom-paged-attn", action="store_true", help="Use custom paged attention"
+    )
     args = parser.parse_args()
     print(args)
 

benchmarks/kernels/benchmark_quant.py

@@ -10,15 +10,17 @@ from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
 
 
 @torch.inference_mode()
-def main(num_tokens: int,
-         hidden_size: int,
-         static_scale: bool,
-         quant_dtype: torch.dtype,
-         dtype: torch.dtype,
-         seed: int = 0,
-         do_profile: bool = False,
-         num_warmup_iters: int = 5,
-         num_iters: int = 100) -> None:
+def main(
+    num_tokens: int,
+    hidden_size: int,
+    static_scale: bool,
+    quant_dtype: torch.dtype,
+    dtype: torch.dtype,
+    seed: int = 0,
+    do_profile: bool = False,
+    num_warmup_iters: int = 5,
+    num_iters: int = 100,
+) -> None:
     current_platform.seed_everything(seed)
     torch.set_default_device("cuda")
 
@@ -56,7 +58,7 @@ def main(num_tokens: int,
     print(f"Kernel running time: {latency * 1000000:.3f} us")
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
 
     def to_torch_dtype(dt):
         if dt == "int8":
@@ -66,37 +68,40 @@ if __name__ == '__main__':
         raise ValueError(f"Unsupported dtype: {dt}")
 
     parser = FlexibleArgumentParser(
-        description="Benchmark the quantization (fp8 or int8) kernel.")
+        description="Benchmark the quantization (fp8 or int8) kernel."
+    )
     parser.add_argument("--num-tokens", type=int, default=4096)
     parser.add_argument("--hidden-size", type=int, default=8192)
     parser.add_argument("--static-scale", action="store_true")
-    parser.add_argument("--quant-dtype",
-                        type=str,
-                        choices=["fp8", "int8"],
-                        default="int8")
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["half", "bfloat16", "float"],
-                        default="half")
+    parser.add_argument(
+        "--quant-dtype", type=str, choices=["fp8", "int8"], default="int8"
+    )
+    parser.add_argument(
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
+    )
 
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--profile", action="store_true")
     parser.add_argument("--num-warmup-iters", type=int, default=5)
-    parser.add_argument("--num-iters",
-                        type=int,
-                        default=100,
-                        help="Number of benchmark iterations. "
-                        "If --profile is set, this number is ignored")
+    parser.add_argument(
+        "--num-iters",
+        type=int,
+        default=100,
+        help="Number of benchmark iterations. "
+        "If --profile is set, this number is ignored",
+    )
 
     args = parser.parse_args()
     print(args)
 
-    main(num_tokens=args.num_tokens,
-         hidden_size=args.hidden_size,
-         static_scale=args.static_scale,
-         quant_dtype=to_torch_dtype(args.quant_dtype),
-         dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
-         seed=args.seed,
-         do_profile=args.profile,
-         num_warmup_iters=args.num_warmup_iters,
-         num_iters=args.num_iters)
+    main(
+        num_tokens=args.num_tokens,
+        hidden_size=args.hidden_size,
+        static_scale=args.static_scale,
+        quant_dtype=to_torch_dtype(args.quant_dtype),
+        dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+        seed=args.seed,
+        do_profile=args.profile,
+        num_warmup_iters=args.num_warmup_iters,
+        num_iters=args.num_iters,
+    )

benchmarks/kernels/benchmark_rmsnorm.py

@@ -4,15 +4,14 @@ import itertools
 from typing import Optional, Union
 
 import torch
-import triton
 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))
@@ -114,23 +113,19 @@ def rmsnorm_vllm(
 
 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")
+    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)
+        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)
+        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)
+        x.clone(), weight, residual.clone() if residual is not None else None
+    )
 
     if use_residual:
         output_naive = output_naive[0]
@@ -141,9 +136,9 @@ def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):
     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):
+    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")
@@ -152,12 +147,10 @@ def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):
 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))
+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"],
@@ -167,19 +160,15 @@ def get_benchmark(use_residual):
             line_names=["HuggingFace", "FlashInfer", "vLLM"],
             styles=[("blue", "-"), ("green", "-"), ("red", "-")],
             ylabel="us",
-            plot_name=
-            f"rmsnorm-perf-{'with' if use_residual else 'without'}-residual",
+            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")
+        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
 
@@ -240,9 +229,9 @@ if __name__ == "__main__":
         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(
+        "--use-residual", action="store_true", help="Whether to use residual connection"
+    )
     parser.add_argument(
         "--save-path",
         type=str,
@@ -253,10 +242,12 @@ if __name__ == "__main__":
     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)
+    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)

benchmarks/kernels/benchmark_rope.py

@@ -6,8 +6,7 @@ from typing import Optional
 import nvtx
 import torch
 
-from vllm.model_executor.layers.rotary_embedding import (RotaryEmbedding,
-                                                         get_rope)
+from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope
 from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser
 
@@ -32,40 +31,49 @@ def benchmark_rope_kernels_multi_lora(
     # 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, {
-                                "rope_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] = []
     for scaling_factor in scaling_factors:
         non_batched_ropes.append(
-            get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
-                     {
-                         "rope_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
@@ -86,27 +94,28 @@ 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_w8a8_block_fp8.py

@@ -14,14 +14,16 @@ import tqdm
 import triton
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    _w8a8_block_fp8_matmul)
+    _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."
+assert current_platform.is_cuda(), (
+    "Only support tune w8a8 block fp8 kernel on CUDA device."
+)
 
 DTYPE_MAP = {
     "float32": torch.float32,
@@ -40,7 +42,7 @@ def w8a8_block_matmul(
     config: dict[str, Any],
     output_dtype: torch.dtype = torch.float16,
 ) -> torch.Tensor:
-    """This function performs matrix multiplication with 
+    """This function performs matrix multiplication with
     block-wise quantization.
 
     It takes two input tensors `A` and `B` with scales `As` and `Bs`.
@@ -51,7 +53,7 @@ def w8a8_block_matmul(
         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. 
+        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.
 
@@ -71,18 +73,18 @@ def w8a8_block_matmul(
     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_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"]), )
+        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.")
+        raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.")
 
     kernel[grid](
         A,
@@ -119,14 +121,16 @@ def get_configs_compute_bound():
                 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,
-                            })
+                            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
 
 
@@ -165,15 +169,9 @@ def get_weight_shapes(tp_size):
     return weight_shapes
 
 
-def benchmark_config(A,
-                     B,
-                     As,
-                     Bs,
-                     block_size,
-                     config,
-                     out_dtype=torch.float16,
-                     num_iters=10):
-
+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)
 
@@ -206,26 +204,26 @@ def tune(M, N, K, block_size, out_dtype, search_space, input_type):
         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)
+            (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)
+            (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.")
+        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)
+    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")
@@ -267,7 +265,8 @@ def save_configs(
     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")
+        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}...")
@@ -295,8 +294,7 @@ def tune_on_gpu(args_dict):
 
     search_space = get_configs_compute_bound()
     search_space = [
-        config for config in search_space
-        if block_k % config["BLOCK_SIZE_K"] == 0
+        config for config in search_space if block_k % config["BLOCK_SIZE_K"] == 0
     ]
 
     start = time.time()
@@ -312,15 +310,11 @@ def tune_on_gpu(args_dict):
                 out_dtype,
                 search_space,
                 input_type,
-            ) for batch_size in tqdm(batch_sizes,
-                                     desc=f"GPU {gpu_id} - Batch sizes")
+            )
+            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)
+        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")
@@ -376,13 +370,14 @@ def main(args):
 
     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,
-        })
+        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:
@@ -398,13 +393,11 @@ 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)
+        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("--input-type", type=str, choices=["fp8"], default="fp8")
     parser.add_argument(
         "--out-dtype",
         type=str,

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -6,13 +6,15 @@ import time
 # Import DeepGEMM functions
 import deep_gemm
 import torch
-import triton
 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)
+    per_token_group_quant_fp8,
+    w8a8_block_fp8_matmul,
+)
+from vllm.triton_utils import triton
 
 
 # Copied from

benchmarks/kernels/graph_machete_bench.py

@@ -14,13 +14,14 @@ 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:
+    with open(args.filename, "rb") as f:
         data = pickle.load(f)
         raw_results: list[TMeasurement] = data["results"]
 
@@ -38,11 +39,7 @@ 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))
@@ -50,14 +47,16 @@ 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

@@ -23,6 +23,7 @@ class ArgPool:
     For every invocation during a benchmarking run, it will choose a
     different value from the list.
     """
+
     values: Iterable[Any]
 
     def __getitem__(self, index):
@@ -30,9 +31,7 @@ class ArgPool:
 
 
 class Bench:
-
     class ArgsIterator:
-
         def __init__(self, args_list, kwargs_list):
             assert len(args_list) == len(kwargs_list)
             self.args_list = args_list
@@ -53,10 +52,16 @@ class Bench:
         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):
-
+    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
@@ -67,10 +72,8 @@ class Bench:
         # 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)
+        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
@@ -100,16 +103,13 @@ class Bench:
 
         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]
-            ])
+            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)
-            ]
+            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]
@@ -142,7 +142,7 @@ class Bench:
 
     def run_cudagrah(self) -> TMeasurement:
         assert self.use_cuda_graph
-        globals = {'g': self.g}
+        globals = {"g": self.g}
 
         return TBenchmark.Timer(
             stmt="g.replay()",
@@ -162,15 +162,15 @@ class Bench:
 
         has_arg_pool = self.args_iterator.n_args > 1
         if has_arg_pool:
-            setup = '''
+            setup = """
                     args_iterator.reset()
                     args_it = args_iterator.__next__()
-                    '''
-            stmt = '''
+                    """
+            stmt = """
                     args, kwargs = next(args_it)
                     fn(*args, **kwargs)
-                    '''
-            globals = {'fn': self.fn, 'args_iterator': self.args_iterator}
+                    """
+            globals = {"fn": self.fn, "args_iterator": self.args_iterator}
         else:
             # no arg pool. Just use the args and kwargs directly
             self.args_iterator.reset()
@@ -178,10 +178,10 @@ class Bench:
             args, kwargs = next(args_it)
 
             setup = ""
-            stmt = '''
+            stmt = """
                     fn(*args, **kwargs)
-                   '''
-            globals = {'fn': self.fn, 'args': args, 'kwargs': kwargs}
+                   """
+            globals = {"fn": self.fn, "args": args, "kwargs": kwargs}
 
         return TBenchmark.Timer(
             stmt=stmt,

benchmarks/overheads/benchmark_hashing.py

@@ -7,9 +7,8 @@ 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):
@@ -30,32 +29,35 @@ 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')
+        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
+exclude = [
+    # External file, leaving license intact
+    "examples/other/fp8/quantizer/quantize.py",
+    "vllm/vllm_flash_attn/flash_attn_interface.pyi"
+]
+
+[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
 
-# Define the model to use
-MODEL=${1:-"Qwen/Qwen2.5-7B-Instruct"}
-
-# Define the backend to use
-BACKEND=${2:-"vllm"}
-
-# Define the dataset to use
-DATASET=${3:-"xgrammar_bench"}
-
-# Define the guided decoding backend
-GUIDED_BACKEND=${4:-"xgrammar"}
-
+# 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=${5:-"$SCRIPT_DIR/structured_output_benchmark_results"}
-
-GUIDED_RATIO=${6:-0.5}
+OUTPUT_DIR=${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=(70 60 50 25 20 15 10)
+QPS_VALUES=(25 20 15 10 5 1)
 
 # Common parameters
 COMMON_PARAMS="--backend $BACKEND \
                --model $MODEL \
                --dataset $DATASET \
-               --structured-output-backend $GUIDED_BACKEND \
-               --structured-output-ratio $GUIDED_RATIO \
+               --structured-output-ratio $STRUCTURED_OUTPUT_RATIO \
                --save-results \
-               --result-dir $OUTPUT_DIR"
+               --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 "Structured output backend: $GUIDED_BACKEND"
 echo "Results will be saved to: $OUTPUT_DIR"
 echo "----------------------------------------"
 
@@ -48,14 +109,17 @@ for qps in "${QPS_VALUES[@]}"; do
   GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown")
 
   # Construct filename for this run
-  FILENAME="${GUIDED_BACKEND}_${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json"
+  FILENAME="${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" \
-    --tokenizer-mode ${TOKENIZER_MODE:-"auto"} \
-    --port ${PORT:-8000}
+    --num-prompts $NUM_PROMPTS
 
   echo "Completed benchmark with QPS: $qps"
   echo "----------------------------------------"

cmake/cpu_extension.cmake

@@ -167,6 +167,33 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
 
     FetchContent_MakeAvailable(oneDNN)
     
+    list(APPEND LIBS dnnl)
+elseif(POWER10_FOUND)
+    FetchContent_Declare(
+        oneDNN
+        GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
+        GIT_TAG v3.7.2
+        GIT_PROGRESS TRUE
+        GIT_SHALLOW TRUE
+    )
+
+    set(ONEDNN_LIBRARY_TYPE "STATIC")
+    set(ONEDNN_BUILD_DOC "OFF")
+    set(ONEDNN_BUILD_EXAMPLES "OFF")
+    set(ONEDNN_BUILD_TESTS "OFF")
+    set(ONEDNN_ENABLE_WORKLOAD "INFERENCE")
+    set(ONEDNN_ENABLE_PRIMITIVE "MATMUL;REORDER")
+    set(ONEDNN_BUILD_GRAPH "OFF")
+    set(ONEDNN_ENABLE_JIT_PROFILING "OFF")
+    set(ONEDNN_ENABLE_ITT_TASKS "OFF")
+    set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
+    set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
+    set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
+
+    set(DNNL_CPU_RUNTIME "OMP")
+
+    FetchContent_MakeAvailable(oneDNN)
+
     list(APPEND LIBS dnnl)
 endif()
 
@@ -197,6 +224,10 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         "csrc/cpu/quant.cpp"
         "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
+elseif(POWER10_FOUND)
+    set(VLLM_EXT_SRC
+        "csrc/cpu/quant.cpp"
+        ${VLLM_EXT_SRC})
 endif()
 
 #
@@ -214,4 +245,4 @@ define_gpu_extension_target(
     WITH_SOABI
 )
 
-message(STATUS "Enabling C extension.")
\ No newline at end of file
+message(STATUS "Enabling C extension.")

cmake/utils.cmake

@@ -229,11 +229,26 @@ macro(set_gencode_flags_for_srcs)
                         "${multiValueArgs}" ${ARGN} )
 
   foreach(_ARCH ${arg_CUDA_ARCHS})
-    string(REPLACE "." "" _ARCH "${_ARCH}")
-    set_gencode_flag_for_srcs(
-      SRCS ${arg_SRCS}
-      ARCH "compute_${_ARCH}"
-      CODE "sm_${_ARCH}")
+    # handle +PTX suffix: generate both sm and ptx codes if requested
+    string(FIND "${_ARCH}" "+PTX" _HAS_PTX)
+    if(NOT _HAS_PTX EQUAL -1)
+      string(REPLACE "+PTX" "" _BASE_ARCH "${_ARCH}")
+      string(REPLACE "." "" _STRIPPED_ARCH "${_BASE_ARCH}")
+      set_gencode_flag_for_srcs(
+        SRCS ${arg_SRCS}
+        ARCH "compute_${_STRIPPED_ARCH}"
+        CODE "sm_${_STRIPPED_ARCH}")
+      set_gencode_flag_for_srcs(
+        SRCS ${arg_SRCS}
+        ARCH "compute_${_STRIPPED_ARCH}"
+        CODE "compute_${_STRIPPED_ARCH}")
+    else()
+      string(REPLACE "." "" _STRIPPED_ARCH "${_ARCH}")
+      set_gencode_flag_for_srcs(
+        SRCS ${arg_SRCS}
+        ARCH "compute_${_STRIPPED_ARCH}"
+        CODE "sm_${_STRIPPED_ARCH}")
+    endif()
   endforeach()
 
   if (${arg_BUILD_PTX_FOR_ARCH})
@@ -252,7 +267,10 @@ endmacro()
 #
 # For the given `SRC_CUDA_ARCHS` list of gencode versions in the form 
 #  `<major>.<minor>[letter]` compute the "loose intersection" with the 
-#  `TGT_CUDA_ARCHS` list of gencodes. 
+#  `TGT_CUDA_ARCHS` list of gencodes. We also support the `+PTX` suffix in
+#  `SRC_CUDA_ARCHS` which indicates that the PTX code should be built when there
+#  is a CUDA_ARCH in `TGT_CUDA_ARCHS` that is equal to or larger than the
+#  architecture in `SRC_CUDA_ARCHS`.
 # The loose intersection is defined as:
 #   { max{ x \in tgt | x <= y } | y \in src, { x \in tgt | x <= y } != {} }
 #  where `<=` is the version comparison operator.
@@ -269,44 +287,63 @@ endmacro()
 #   cuda_archs_loose_intersection(OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_ARCHS)
 #   OUT_CUDA_ARCHS="8.0;8.6;9.0;9.0a"
 #
+# Example With PTX:
+#   SRC_CUDA_ARCHS="8.0+PTX"
+#   TGT_CUDA_ARCHS="9.0"
+#   cuda_archs_loose_intersection(OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_ARCHS)
+#   OUT_CUDA_ARCHS="8.0+PTX"
+#
 function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_ARCHS)
-  list(REMOVE_DUPLICATES SRC_CUDA_ARCHS)
-  set(TGT_CUDA_ARCHS_ ${TGT_CUDA_ARCHS})
+  set(_SRC_CUDA_ARCHS "${SRC_CUDA_ARCHS}")
+  set(_TGT_CUDA_ARCHS ${TGT_CUDA_ARCHS})
+
+  # handle +PTX suffix: separate base arch for matching, record PTX requests
+  set(_PTX_ARCHS)
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "\\+PTX$")
+      string(REPLACE "+PTX" "" _base "${_arch}")
+      list(APPEND _PTX_ARCHS "${_base}")
+      list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+      list(APPEND _SRC_CUDA_ARCHS "${_base}")
+    endif()
+  endforeach()
+  list(REMOVE_DUPLICATES _PTX_ARCHS)
+  list(REMOVE_DUPLICATES _SRC_CUDA_ARCHS)
 
   # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
   # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
-  if ("9.0a" IN_LIST SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM SRC_CUDA_ARCHS "9.0a")
-    if ("9.0" IN_LIST TGT_CUDA_ARCHS_)
-      list(REMOVE_ITEM TGT_CUDA_ARCHS_ "9.0")
+  if ("9.0a" IN_LIST _SRC_CUDA_ARCHS)
+    list(REMOVE_ITEM _SRC_CUDA_ARCHS "9.0a")
+    if ("9.0" IN_LIST TGT_CUDA_ARCHS)
+      list(REMOVE_ITEM _TGT_CUDA_ARCHS "9.0")
       set(_CUDA_ARCHS "9.0a")
     endif()
   endif()
 
-  if ("10.0a" IN_LIST SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM SRC_CUDA_ARCHS "10.0a")
+  if ("10.0a" IN_LIST _SRC_CUDA_ARCHS)
+    list(REMOVE_ITEM _SRC_CUDA_ARCHS "10.0a")
     if ("10.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM TGT_CUDA_ARCHS_ "10.0")
+      list(REMOVE_ITEM _TGT_CUDA_ARCHS "10.0")
       set(_CUDA_ARCHS "10.0a")
     endif()
   endif()
 
-  list(SORT SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
+  list(SORT _SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
   # for each ARCH in TGT_CUDA_ARCHS find the highest arch in SRC_CUDA_ARCHS that
   # is less or equal to ARCH (but has the same major version since SASS binary
   # compatibility is only forward compatible within the same major version).
-  foreach(_ARCH ${TGT_CUDA_ARCHS_})
+  foreach(_ARCH ${_TGT_CUDA_ARCHS})
     set(_TMP_ARCH)
     # Extract the major version of the target arch
     string(REGEX REPLACE "^([0-9]+)\\..*$" "\\1" TGT_ARCH_MAJOR "${_ARCH}")
-    foreach(_SRC_ARCH ${SRC_CUDA_ARCHS})
+    foreach(_SRC_ARCH ${_SRC_CUDA_ARCHS})
       # Extract the major version of the source arch
       string(REGEX REPLACE "^([0-9]+)\\..*$" "\\1" SRC_ARCH_MAJOR "${_SRC_ARCH}")
-      # Check major-version match AND version-less-or-equal
+      # Check version-less-or-equal, and allow PTX arches to match across majors
       if (_SRC_ARCH VERSION_LESS_EQUAL _ARCH)
-        if (SRC_ARCH_MAJOR STREQUAL TGT_ARCH_MAJOR)
+        if (_SRC_ARCH IN_LIST _PTX_ARCHS OR SRC_ARCH_MAJOR STREQUAL TGT_ARCH_MAJOR)
           set(_TMP_ARCH "${_SRC_ARCH}")
         endif()
       else()
@@ -322,6 +359,18 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   endforeach()
 
   list(REMOVE_DUPLICATES _CUDA_ARCHS)
+  
+  # reapply +PTX suffix to architectures that requested PTX
+  set(_FINAL_ARCHS)
+  foreach(_arch ${_CUDA_ARCHS})
+    if(_arch IN_LIST _PTX_ARCHS)
+      list(APPEND _FINAL_ARCHS "${_arch}+PTX")
+    else()
+      list(APPEND _FINAL_ARCHS "${_arch}")
+    endif()
+  endforeach()
+  set(_CUDA_ARCHS ${_FINAL_ARCHS})
+  
   set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE)
 endfunction()
 

csrc/activation_kernels.cu

@@ -70,6 +70,9 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   int64_t num_tokens = input.numel() / input.size(-1);                   \
   dim3 grid(num_tokens);                                                 \
   dim3 block(std::min(d, 1024));                                         \
+  if (num_tokens == 0) {                                                 \
+    return;                                                              \
+  }                                                                      \
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
   VLLM_DISPATCH_FLOATING_TYPES(                                          \

csrc/attention/attention_kernels.cuh

@@ -17,660 +17,660 @@
  * limitations under the License.
  */
 
- #include <torch/all.h>
- #include <ATen/cuda/CUDAContext.h>
- #include <c10/cuda/CUDAGuard.h>
- #include <algorithm>
- 
- #include "attention_dtypes.h"
- #include "attention_utils.cuh"
- 
- #ifdef USE_ROCM
-   #include <hip/hip_bf16.h>
-   #include "../quantization/fp8/amd/quant_utils.cuh"
- typedef __hip_bfloat16 __nv_bfloat16;
- #else
-   #include "../quantization/fp8/nvidia/quant_utils.cuh"
- #endif
- 
- #ifndef USE_ROCM
-   #define WARP_SIZE 32
- #else
-   #define WARP_SIZE warpSize
- #endif
- 
- #define MAX(a, b) ((a) > (b) ? (a) : (b))
- #define MIN(a, b) ((a) < (b) ? (a) : (b))
- #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
- 
- namespace vllm {
- 
- // Utility function for attention softmax.
- template <int NUM_WARPS>
- inline __device__ float block_sum(float* red_smem, float sum) {
-   // Decompose the thread index into warp / lane.
-   int warp = threadIdx.x / WARP_SIZE;
-   int lane = threadIdx.x % WARP_SIZE;
- 
-   // Compute the sum per warp.
- #pragma unroll
-   for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
-     sum += VLLM_SHFL_XOR_SYNC(sum, mask);
-   }
- 
-   // Warp leaders store the data to shared memory.
-   if (lane == 0) {
-     red_smem[warp] = sum;
-   }
- 
-   // Make sure the data is in shared memory.
-   __syncthreads();
- 
-   // The warps compute the final sums.
-   if (lane < NUM_WARPS) {
-     sum = red_smem[lane];
-   }
- 
-   // Parallel reduction inside the warp.
- #pragma unroll
-   for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
-     sum += VLLM_SHFL_XOR_SYNC(sum, mask);
-   }
- 
-   // Broadcast to other threads.
-   return VLLM_SHFL_SYNC(sum, 0);
- }
- 
- // TODO(woosuk): Merge the last two dimensions of the grid.
- // Grid: (num_heads, num_seqs, max_num_partitions).
- template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
-           int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
-           bool IS_BLOCK_SPARSE,
-           int PARTITION_SIZE = 0>  // Zero means no partitioning.
- __device__ void paged_attention_kernel(
-     float* __restrict__ exp_sums,  // [num_seqs, num_heads, max_num_partitions]
-     float* __restrict__ max_logits,  // [num_seqs, num_heads,
-                                      // max_num_partitions]
-     scalar_t* __restrict__ out,  // [num_seqs, num_heads, max_num_partitions,
-                                  // head_size]
-     const scalar_t* __restrict__ q,       // [num_seqs, num_heads, head_size]
-     const cache_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
-                                           // head_size/x, block_size, x]
-     const cache_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
-                                           // head_size, block_size]
-     const int num_kv_heads,               // [num_heads]
-     const float scale,
-     const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
-     const int* __restrict__ seq_lens,      // [num_seqs]
-     const int max_num_blocks_per_seq,
-     const float* __restrict__ alibi_slopes,  // [num_heads]
-     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-     const float* k_scale, const float* v_scale, const int tp_rank,
-     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
-   const int seq_idx = blockIdx.y;
-   const int partition_idx = blockIdx.z;
-   const int max_num_partitions = gridDim.z;
-   constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
-   const int seq_len = seq_lens[seq_idx];
-   if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= seq_len) {
-     // No work to do. Terminate the thread block.
-     return;
-   }
- 
-   const int num_seq_blocks = DIVIDE_ROUND_UP(seq_len, BLOCK_SIZE);
-   const int num_blocks_per_partition =
-       USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_seq_blocks;
- 
-   // [start_block_idx, end_block_idx) is the range of blocks to process.
-   const int start_block_idx =
-       USE_PARTITIONING ? partition_idx * num_blocks_per_partition : 0;
-   const int end_block_idx =
-       MIN(start_block_idx + num_blocks_per_partition, num_seq_blocks);
-   const int num_blocks = end_block_idx - start_block_idx;
- 
-   // [start_token_idx, end_token_idx) is the range of tokens to process.
-   const int start_token_idx = start_block_idx * BLOCK_SIZE;
-   const int end_token_idx =
-       MIN(start_token_idx + num_blocks * BLOCK_SIZE, seq_len);
-   const int num_tokens = end_token_idx - start_token_idx;
- 
-   constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
-   constexpr int NUM_THREAD_GROUPS =
-       NUM_THREADS / THREAD_GROUP_SIZE;  // Note: This assumes THREAD_GROUP_SIZE
-                                         // divides NUM_THREADS
-   assert(NUM_THREADS % THREAD_GROUP_SIZE == 0);
-   constexpr int NUM_TOKENS_PER_THREAD_GROUP =
-       DIVIDE_ROUND_UP(BLOCK_SIZE, WARP_SIZE);
-   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-   const int thread_idx = threadIdx.x;
-   const int warp_idx = thread_idx / WARP_SIZE;
-   const int lane = thread_idx % WARP_SIZE;
- 
-   const int head_idx = blockIdx.x;
-   const int num_heads = gridDim.x;
-   const int num_queries_per_kv = num_heads / num_kv_heads;
-   const int kv_head_idx = head_idx / num_queries_per_kv;
-   const float alibi_slope =
-       alibi_slopes == nullptr ? 0.f : alibi_slopes[head_idx];
- 
-   // A vector type to store a part of a key or a query.
-   // The vector size is configured in such a way that the threads in a thread
-   // group fetch or compute 16 bytes at a time. For example, if the size of a
-   // thread group is 4 and the data type is half, then the vector size is 16 /
-   // (4 * sizeof(half)) == 2.
-   constexpr int VEC_SIZE = MAX(16 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
-   using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-   using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-   using Quant_vec = typename Vec<cache_t, VEC_SIZE>::Type;
- 
-   constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
-   constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
- 
-   const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
-   const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
- 
-   // Load the query to registers.
-   // Each thread in a thread group has a different part of the query.
-   // For example, if the the thread group size is 4, then the first thread in
-   // the group has 0, 4, 8, ... th vectors of the query, and the second thread
-   // has 1, 5, 9, ... th vectors of the query, and so on. NOTE(woosuk): Because
-   // q is split from a qkv tensor, it may not be contiguous.
-   const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
-   __shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
- #pragma unroll
-   for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD;
-        i += NUM_THREAD_GROUPS) {
-     const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
-     q_vecs[thread_group_offset][i] =
-         *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
-   }
-   __syncthreads();  // TODO(naed90): possible speedup if this is replaced with a
-                     // memory wall right before we use q_vecs
- 
-   // Memory planning.
-   extern __shared__ char shared_mem[];
-   // NOTE(woosuk): We use FP32 for the softmax logits for better accuracy.
-   float* logits = reinterpret_cast<float*>(shared_mem);
-   // Workspace for reduction.
-   __shared__ float red_smem[2 * NUM_WARPS];
- 
-   // x == THREAD_GROUP_SIZE * VEC_SIZE
-   // Each thread group fetches x elements from the key at a time.
-   constexpr int x = 16 / sizeof(cache_t);
-   float qk_max = -FLT_MAX;
- 
-   // Iterate over the key blocks.
-   // Each warp fetches a block of keys for each iteration.
-   // Each thread group in a warp fetches a key from the block, and computes
-   // dot product with the query.
-   const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
- 
-   // blocksparse specific vars
-   int bs_block_offset;
-   int q_bs_block_id;
-   if constexpr (IS_BLOCK_SPARSE) {
-     // const int num_blocksparse_blocks = DIVIDE_ROUND_UP(seq_len,
-     // blocksparse_block_size);
-     q_bs_block_id = (seq_len - 1) / blocksparse_block_size;
-     if (blocksparse_head_sliding_step >= 0)
-       // sliding on q heads
-       bs_block_offset =
-           (tp_rank * num_heads + head_idx) * blocksparse_head_sliding_step + 1;
-     else
-       // sliding on kv heads
-       bs_block_offset = (tp_rank * num_kv_heads + kv_head_idx) *
-                             (-blocksparse_head_sliding_step) +
-                         1;
-   }
- 
-   for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
-        block_idx += NUM_WARPS) {
-     // NOTE(woosuk): The block number is stored in int32. However, we cast it to
-     // int64 because int32 can lead to overflow when this variable is multiplied
-     // by large numbers (e.g., kv_block_stride).
-     // For blocksparse attention: skip computation on blocks that are not
-     // attended
-     if constexpr (IS_BLOCK_SPARSE) {
-       const int k_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
-       const bool is_remote =
-           ((k_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0);
-       const bool is_local =
-           (k_bs_block_id > q_bs_block_id - blocksparse_local_blocks);
-       if (!is_remote && !is_local) {
-         for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
-           const int physical_block_offset =
-               (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
-           const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
- 
-           if (thread_group_offset == 0) {
-             // NOTE(linxihui): assign very large number to skipped tokens to
-             // avoid contribution to the sumexp softmax normalizer. This will
-             // not be used at computing sum(softmax*v) as the blocks will be
-             // skipped.
-             logits[token_idx - start_token_idx] = -FLT_MAX;
-           }
-         }
-         continue;
-       }
-     }
-     const int64_t physical_block_number =
-         static_cast<int64_t>(block_table[block_idx]);
- 
-     // Load a key to registers.
-     // Each thread in a thread group has a different part of the key.
-     // For example, if the the thread group size is 4, then the first thread in
-     // the group has 0, 4, 8, ... th vectors of the key, and the second thread
-     // has 1, 5, 9, ... th vectors of the key, and so on.
-     for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
-       const int physical_block_offset =
-           (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
-       const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
-       K_vec k_vecs[NUM_VECS_PER_THREAD];
- 
- #pragma unroll
-       for (int j = 0; j < NUM_VECS_PER_THREAD; j++) {
-         const cache_t* k_ptr =
-             k_cache + physical_block_number * kv_block_stride +
-             kv_head_idx * kv_head_stride + physical_block_offset * x;
-         const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
-         const int offset1 = (vec_idx * VEC_SIZE) / x;
-         const int offset2 = (vec_idx * VEC_SIZE) % x;
- 
-         if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
-           k_vecs[j] = *reinterpret_cast<const K_vec*>(
-               k_ptr + offset1 * BLOCK_SIZE * x + offset2);
-         } else {
-           // Vector conversion from Quant_vec to K_vec.
-           Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
-               k_ptr + offset1 * BLOCK_SIZE * x + offset2);
-           k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
-               k_vec_quant, *k_scale);
-         }
-       }
- 
-       // Compute dot product.
-       // This includes a reduction across the threads in the same thread group.
-       float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(
-                              q_vecs[thread_group_offset], k_vecs);
-       // Add the ALiBi bias if slopes are given.
-       qk += (alibi_slope != 0) ? alibi_slope * (token_idx - seq_len + 1) : 0;
- 
-       if (thread_group_offset == 0) {
-         // Store the partial reductions to shared memory.
-         // NOTE(woosuk): It is required to zero out the masked logits.
-         const bool mask = token_idx >= seq_len;
-         logits[token_idx - start_token_idx] = mask ? 0.f : qk;
-         // Update the max value.
-         qk_max = mask ? qk_max : fmaxf(qk_max, qk);
-       }
-     }
-   }
- 
-   // Perform reduction across the threads in the same warp to get the
-   // max qk value for each "warp" (not across the thread block yet).
-   // The 0-th thread of each thread group already has its max qk value.
- #pragma unroll
-   for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
-     qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
-   }
-   if (lane == 0) {
-     red_smem[warp_idx] = qk_max;
-   }
-   __syncthreads();
- 
-   // TODO(woosuk): Refactor this part.
-   // Get the max qk value for the sequence.
-   qk_max = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
- #pragma unroll
-   for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
-     qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
-   }
-   // Broadcast the max qk value to all threads.
-   qk_max = VLLM_SHFL_SYNC(qk_max, 0);
- 
-   // Get the sum of the exp values.
-   float exp_sum = 0.f;
-   for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
-     float val = __expf(logits[i] - qk_max);
-     logits[i] = val;
-     exp_sum += val;
-   }
-   exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
- 
-   // Compute softmax.
-   const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
-   for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
-     logits[i] *= inv_sum;
-   }
-   __syncthreads();
- 
-   // If partitioning is enabled, store the max logit and exp_sum.
-   if (USE_PARTITIONING && thread_idx == 0) {
-     float* max_logits_ptr = max_logits +
-                             seq_idx * num_heads * max_num_partitions +
-                             head_idx * max_num_partitions + partition_idx;
-     *max_logits_ptr = qk_max;
-     float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions +
-                           head_idx * max_num_partitions + partition_idx;
-     *exp_sums_ptr = exp_sum;
-   }
- 
-   // Each thread will fetch 16 bytes from the value cache at a time.
-   constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
-   using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
-   using L_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
-   using V_quant_vec = typename Vec<cache_t, V_VEC_SIZE>::Type;
-   using Float_L_vec = typename FloatVec<L_vec>::Type;
- 
-   constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
-   constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
-   constexpr int NUM_ROWS_PER_THREAD =
-       DIVIDE_ROUND_UP(HEAD_SIZE, NUM_ROWS_PER_ITER);
- 
-   // NOTE(woosuk): We use FP32 for the accumulator for better accuracy.
-   float accs[NUM_ROWS_PER_THREAD];
- #pragma unroll
-   for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-     accs[i] = 0.f;
-   }
- 
-   scalar_t zero_value;
-   zero(zero_value);
-   for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
-        block_idx += NUM_WARPS) {
-     // NOTE(woosuk): The block number is stored in int32. However, we cast it to
-     // int64 because int32 can lead to overflow when this variable is multiplied
-     // by large numbers (e.g., kv_block_stride).
-     // For blocksparse attention: skip computation on blocks that are not
-     // attended
-     if constexpr (IS_BLOCK_SPARSE) {
-       int v_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
-       if (!((v_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0) &&
-           !((v_bs_block_id > q_bs_block_id - blocksparse_local_blocks))) {
-         continue;
-       }
-     }
-     const int64_t physical_block_number =
-         static_cast<int64_t>(block_table[block_idx]);
-     const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
-     const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
-     L_vec logits_vec;
-     from_float(logits_vec, *reinterpret_cast<Float_L_vec*>(logits + token_idx -
-                                                            start_token_idx));
- 
-     const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride +
-                            kv_head_idx * kv_head_stride;
- #pragma unroll
-     for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-       const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-       if (row_idx < HEAD_SIZE) {
-         const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
-         V_vec v_vec;
- 
-         if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
-           v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
-         } else {
-           V_quant_vec v_quant_vec =
-               *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
-           // Vector conversion from V_quant_vec to V_vec.
-           v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
-                                                                     *v_scale);
-         }
-         if (block_idx == num_seq_blocks - 1) {
-           // NOTE(woosuk): When v_vec contains the tokens that are out of the
-           // context, we should explicitly zero out the values since they may
-           // contain NaNs. See
-           // https://github.com/vllm-project/vllm/issues/641#issuecomment-1682544472
-           scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
- #pragma unroll
-           for (int j = 0; j < V_VEC_SIZE; j++) {
-             v_vec_ptr[j] = token_idx + j < seq_len ? v_vec_ptr[j] : zero_value;
-           }
-         }
-         accs[i] += dot(logits_vec, v_vec);
-       }
-     }
-   }
- 
-   // Perform reduction within each warp.
- #pragma unroll
-   for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-     float acc = accs[i];
- #pragma unroll
-     for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
-       acc += VLLM_SHFL_XOR_SYNC(acc, mask);
-     }
-     accs[i] = acc;
-   }
- 
-   // NOTE(woosuk): A barrier is required because the shared memory space for
-   // logits is reused for the output.
-   __syncthreads();
- 
-   // Perform reduction across warps.
-   float* out_smem = reinterpret_cast<float*>(shared_mem);
- #pragma unroll
-   for (int i = NUM_WARPS; i > 1; i /= 2) {
-     int mid = i / 2;
-     // Upper warps write to shared memory.
-     if (warp_idx >= mid && warp_idx < i) {
-       float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
- #pragma unroll
-       for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-         const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-         if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-           dst[row_idx] = accs[i];
-         }
-       }
-     }
-     __syncthreads();
- 
-     // Lower warps update the output.
-     if (warp_idx < mid) {
-       const float* src = &out_smem[warp_idx * HEAD_SIZE];
- #pragma unroll
-       for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-         const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-         if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-           accs[i] += src[row_idx];
-         }
-       }
-     }
-     __syncthreads();
-   }
- 
-   // Write the final output.
-   if (warp_idx == 0) {
-     scalar_t* out_ptr =
-         out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
-         head_idx * max_num_partitions * HEAD_SIZE + partition_idx * HEAD_SIZE;
- #pragma unroll
-     for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
-       const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
-       if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
-         from_float(*(out_ptr + row_idx), accs[i]);
-       }
-     }
-   }
- }
- 
- // Grid: (num_heads, num_seqs, 1).
- template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
-           int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
-           bool IS_BLOCK_SPARSE>
- __global__ void paged_attention_v1_kernel(
-     scalar_t* __restrict__ out,           // [num_seqs, num_heads, head_size]
-     const scalar_t* __restrict__ q,       // [num_seqs, num_heads, head_size]
-     const cache_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
-                                           // head_size/x, block_size, x]
-     const cache_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
-                                           // head_size, block_size]
-     const int num_kv_heads,               // [num_heads]
-     const float scale,
-     const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
-     const int* __restrict__ seq_lens,      // [num_seqs]
-     const int max_num_blocks_per_seq,
-     const float* __restrict__ alibi_slopes,  // [num_heads]
-     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-     const float* k_scale, const float* v_scale, const int tp_rank,
-     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
-   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
-                          KV_DTYPE, IS_BLOCK_SPARSE>(
-       /* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
-       v_cache, num_kv_heads, scale, block_tables, seq_lens,
-       max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
-       kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
-       blocksparse_vert_stride, blocksparse_block_size,
-       blocksparse_head_sliding_step);
- }
- 
- // Grid: (num_heads, num_seqs, max_num_partitions).
- template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
-           int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
-           bool IS_BLOCK_SPARSE,
-           int PARTITION_SIZE>
- __global__ void paged_attention_v2_kernel(
-     float* __restrict__ exp_sums,  // [num_seqs, num_heads, max_num_partitions]
-     float* __restrict__ max_logits,       // [num_seqs, num_heads,
+#include <torch/all.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <algorithm>
+
+#include "attention_dtypes.h"
+#include "attention_utils.cuh"
+
+#ifdef USE_ROCM
+  #include <hip/hip_bf16.h>
+  #include "../quantization/fp8/amd/quant_utils.cuh"
+typedef __hip_bfloat16 __nv_bfloat16;
+#else
+  #include "../quantization/fp8/nvidia/quant_utils.cuh"
+#endif
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
+
+namespace vllm {
+
+// Utility function for attention softmax.
+template <int NUM_WARPS>
+inline __device__ float block_sum(float* red_smem, float sum) {
+  // Decompose the thread index into warp / lane.
+  int warp = threadIdx.x / WARP_SIZE;
+  int lane = threadIdx.x % WARP_SIZE;
+
+  // Compute the sum per warp.
+#pragma unroll
+  for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
+    sum += VLLM_SHFL_XOR_SYNC(sum, mask);
+  }
+
+  // Warp leaders store the data to shared memory.
+  if (lane == 0) {
+    red_smem[warp] = sum;
+  }
+
+  // Make sure the data is in shared memory.
+  __syncthreads();
+
+  // The warps compute the final sums.
+  if (lane < NUM_WARPS) {
+    sum = red_smem[lane];
+  }
+
+  // Parallel reduction inside the warp.
+#pragma unroll
+  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
+    sum += VLLM_SHFL_XOR_SYNC(sum, mask);
+  }
+
+  // Broadcast to other threads.
+  return VLLM_SHFL_SYNC(sum, 0);
+}
+
+// TODO(woosuk): Merge the last two dimensions of the grid.
+// Grid: (num_heads, num_seqs, max_num_partitions).
+template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
+          int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
+          bool IS_BLOCK_SPARSE,
+          int PARTITION_SIZE = 0>  // Zero means no partitioning.
+__device__ void paged_attention_kernel(
+    float* __restrict__ exp_sums,  // [num_seqs, num_heads, max_num_partitions]
+    float* __restrict__ max_logits,  // [num_seqs, num_heads,
+                                     // max_num_partitions]
+    scalar_t* __restrict__ out,  // [num_seqs, num_heads, max_num_partitions,
+                                 // head_size]
+    const scalar_t* __restrict__ q,       // [num_seqs, num_heads, head_size]
+    const cache_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                          // head_size/x, block_size, x]
+    const cache_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                          // head_size, block_size]
+    const int num_kv_heads,               // [num_heads]
+    const float scale,
+    const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_blocks_per_seq,
+    const float* __restrict__ alibi_slopes,  // [num_heads]
+    const int q_stride, const int kv_block_stride, const int kv_head_stride,
+    const float* k_scale, const float* v_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+  const int seq_idx = blockIdx.y;
+  const int partition_idx = blockIdx.z;
+  const int max_num_partitions = gridDim.z;
+  constexpr bool USE_PARTITIONING = PARTITION_SIZE > 0;
+  const int seq_len = seq_lens[seq_idx];
+  if (USE_PARTITIONING && partition_idx * PARTITION_SIZE >= seq_len) {
+    // No work to do. Terminate the thread block.
+    return;
+  }
+
+  const int num_seq_blocks = DIVIDE_ROUND_UP(seq_len, BLOCK_SIZE);
+  const int num_blocks_per_partition =
+      USE_PARTITIONING ? PARTITION_SIZE / BLOCK_SIZE : num_seq_blocks;
+
+  // [start_block_idx, end_block_idx) is the range of blocks to process.
+  const int start_block_idx =
+      USE_PARTITIONING ? partition_idx * num_blocks_per_partition : 0;
+  const int end_block_idx =
+      MIN(start_block_idx + num_blocks_per_partition, num_seq_blocks);
+  const int num_blocks = end_block_idx - start_block_idx;
+
+  // [start_token_idx, end_token_idx) is the range of tokens to process.
+  const int start_token_idx = start_block_idx * BLOCK_SIZE;
+  const int end_token_idx =
+      MIN(start_token_idx + num_blocks * BLOCK_SIZE, seq_len);
+  const int num_tokens = end_token_idx - start_token_idx;
+
+  constexpr int THREAD_GROUP_SIZE = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  constexpr int NUM_THREAD_GROUPS =
+      NUM_THREADS / THREAD_GROUP_SIZE;  // Note: This assumes THREAD_GROUP_SIZE
+                                        // divides NUM_THREADS
+  assert(NUM_THREADS % THREAD_GROUP_SIZE == 0);
+  constexpr int NUM_TOKENS_PER_THREAD_GROUP =
+      DIVIDE_ROUND_UP(BLOCK_SIZE, WARP_SIZE);
+  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  const int thread_idx = threadIdx.x;
+  const int warp_idx = thread_idx / WARP_SIZE;
+  const int lane = thread_idx % WARP_SIZE;
+
+  const int head_idx = blockIdx.x;
+  const int num_heads = gridDim.x;
+  const int num_queries_per_kv = num_heads / num_kv_heads;
+  const int kv_head_idx = head_idx / num_queries_per_kv;
+  const float alibi_slope =
+      alibi_slopes == nullptr ? 0.f : alibi_slopes[head_idx];
+
+  // A vector type to store a part of a key or a query.
+  // The vector size is configured in such a way that the threads in a thread
+  // group fetch or compute 16 bytes at a time. For example, if the size of a
+  // thread group is 4 and the data type is half, then the vector size is 16 /
+  // (4 * sizeof(half)) == 2.
+  constexpr int VEC_SIZE = MAX(16 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
+  using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
+  using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
+  using Quant_vec = typename Vec<cache_t, VEC_SIZE>::Type;
+
+  constexpr int NUM_ELEMS_PER_THREAD = HEAD_SIZE / THREAD_GROUP_SIZE;
+  constexpr int NUM_VECS_PER_THREAD = NUM_ELEMS_PER_THREAD / VEC_SIZE;
+
+  const int thread_group_idx = thread_idx / THREAD_GROUP_SIZE;
+  const int thread_group_offset = thread_idx % THREAD_GROUP_SIZE;
+
+  // Load the query to registers.
+  // Each thread in a thread group has a different part of the query.
+  // For example, if the thread group size is 4, then the first thread in
+  // the group has 0, 4, 8, ... th vectors of the query, and the second thread
+  // has 1, 5, 9, ... th vectors of the query, and so on. NOTE(woosuk): Because
+  // q is split from a qkv tensor, it may not be contiguous.
+  const scalar_t* q_ptr = q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+  __shared__ Q_vec q_vecs[THREAD_GROUP_SIZE][NUM_VECS_PER_THREAD];
+#pragma unroll
+  for (int i = thread_group_idx; i < NUM_VECS_PER_THREAD;
+       i += NUM_THREAD_GROUPS) {
+    const int vec_idx = thread_group_offset + i * THREAD_GROUP_SIZE;
+    q_vecs[thread_group_offset][i] =
+        *reinterpret_cast<const Q_vec*>(q_ptr + vec_idx * VEC_SIZE);
+  }
+  __syncthreads();  // TODO(naed90): possible speedup if this is replaced with a
+                    // memory wall right before we use q_vecs
+
+  // Memory planning.
+  extern __shared__ char shared_mem[];
+  // NOTE(woosuk): We use FP32 for the softmax logits for better accuracy.
+  float* logits = reinterpret_cast<float*>(shared_mem);
+  // Workspace for reduction.
+  __shared__ float red_smem[2 * NUM_WARPS];
+
+  // x == THREAD_GROUP_SIZE * VEC_SIZE
+  // Each thread group fetches x elements from the key at a time.
+  constexpr int x = 16 / sizeof(cache_t);
+  float qk_max = -FLT_MAX;
+
+  // Iterate over the key blocks.
+  // Each warp fetches a block of keys for each iteration.
+  // Each thread group in a warp fetches a key from the block, and computes
+  // dot product with the query.
+  const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
+
+  // blocksparse specific vars
+  int bs_block_offset;
+  int q_bs_block_id;
+  if constexpr (IS_BLOCK_SPARSE) {
+    // const int num_blocksparse_blocks = DIVIDE_ROUND_UP(seq_len,
+    // blocksparse_block_size);
+    q_bs_block_id = (seq_len - 1) / blocksparse_block_size;
+    if (blocksparse_head_sliding_step >= 0)
+      // sliding on q heads
+      bs_block_offset =
+          (tp_rank * num_heads + head_idx) * blocksparse_head_sliding_step + 1;
+    else
+      // sliding on kv heads
+      bs_block_offset = (tp_rank * num_kv_heads + kv_head_idx) *
+                            (-blocksparse_head_sliding_step) +
+                        1;
+  }
+
+  for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
+       block_idx += NUM_WARPS) {
+    // NOTE(woosuk): The block number is stored in int32. However, we cast it to
+    // int64 because int32 can lead to overflow when this variable is multiplied
+    // by large numbers (e.g., kv_block_stride).
+    // For blocksparse attention: skip computation on blocks that are not
+    // attended
+    if constexpr (IS_BLOCK_SPARSE) {
+      const int k_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
+      const bool is_remote =
+          ((k_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0);
+      const bool is_local =
+          (k_bs_block_id > q_bs_block_id - blocksparse_local_blocks);
+      if (!is_remote && !is_local) {
+        for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
+          const int physical_block_offset =
+              (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
+          const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
+
+          if (thread_group_offset == 0) {
+            // NOTE(linxihui): assign very large number to skipped tokens to
+            // avoid contribution to the sumexp softmax normalizer. This will
+            // not be used at computing sum(softmax*v) as the blocks will be
+            // skipped.
+            logits[token_idx - start_token_idx] = -FLT_MAX;
+          }
+        }
+        continue;
+      }
+    }
+    const int64_t physical_block_number =
+        static_cast<int64_t>(block_table[block_idx]);
+
+    // Load a key to registers.
+    // Each thread in a thread group has a different part of the key.
+    // For example, if the thread group size is 4, then the first thread in
+    // the group has 0, 4, 8, ... th vectors of the key, and the second thread
+    // has 1, 5, 9, ... th vectors of the key, and so on.
+    for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
+      const int physical_block_offset =
+          (thread_group_idx + i * WARP_SIZE) % BLOCK_SIZE;
+      const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
+      K_vec k_vecs[NUM_VECS_PER_THREAD];
+
+#pragma unroll
+      for (int j = 0; j < NUM_VECS_PER_THREAD; j++) {
+        const cache_t* k_ptr =
+            k_cache + physical_block_number * kv_block_stride +
+            kv_head_idx * kv_head_stride + physical_block_offset * x;
+        const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
+        const int offset1 = (vec_idx * VEC_SIZE) / x;
+        const int offset2 = (vec_idx * VEC_SIZE) % x;
+
+        if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
+          k_vecs[j] = *reinterpret_cast<const K_vec*>(
+              k_ptr + offset1 * BLOCK_SIZE * x + offset2);
+        } else {
+          // Vector conversion from Quant_vec to K_vec.
+          Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
+              k_ptr + offset1 * BLOCK_SIZE * x + offset2);
+          k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
+              k_vec_quant, *k_scale);
+        }
+      }
+
+      // Compute dot product.
+      // This includes a reduction across the threads in the same thread group.
+      float qk = scale * Qk_dot<scalar_t, THREAD_GROUP_SIZE>::dot(
+                             q_vecs[thread_group_offset], k_vecs);
+      // Add the ALiBi bias if slopes are given.
+      qk += (alibi_slope != 0) ? alibi_slope * (token_idx - seq_len + 1) : 0;
+
+      if (thread_group_offset == 0) {
+        // Store the partial reductions to shared memory.
+        // NOTE(woosuk): It is required to zero out the masked logits.
+        const bool mask = token_idx >= seq_len;
+        logits[token_idx - start_token_idx] = mask ? 0.f : qk;
+        // Update the max value.
+        qk_max = mask ? qk_max : fmaxf(qk_max, qk);
+      }
+    }
+  }
+
+  // Perform reduction across the threads in the same warp to get the
+  // max qk value for each "warp" (not across the thread block yet).
+  // The 0-th thread of each thread group already has its max qk value.
+#pragma unroll
+  for (int mask = WARP_SIZE / 2; mask >= THREAD_GROUP_SIZE; mask /= 2) {
+    qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
+  }
+  if (lane == 0) {
+    red_smem[warp_idx] = qk_max;
+  }
+  __syncthreads();
+
+  // TODO(woosuk): Refactor this part.
+  // Get the max qk value for the sequence.
+  qk_max = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
+#pragma unroll
+  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
+    qk_max = fmaxf(qk_max, VLLM_SHFL_XOR_SYNC(qk_max, mask));
+  }
+  // Broadcast the max qk value to all threads.
+  qk_max = VLLM_SHFL_SYNC(qk_max, 0);
+
+  // Get the sum of the exp values.
+  float exp_sum = 0.f;
+  for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
+    float val = __expf(logits[i] - qk_max);
+    logits[i] = val;
+    exp_sum += val;
+  }
+  exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], exp_sum);
+
+  // Compute softmax.
+  const float inv_sum = __fdividef(1.f, exp_sum + 1e-6f);
+  for (int i = thread_idx; i < num_tokens; i += NUM_THREADS) {
+    logits[i] *= inv_sum;
+  }
+  __syncthreads();
+
+  // If partitioning is enabled, store the max logit and exp_sum.
+  if (USE_PARTITIONING && thread_idx == 0) {
+    float* max_logits_ptr = max_logits +
+                            seq_idx * num_heads * max_num_partitions +
+                            head_idx * max_num_partitions + partition_idx;
+    *max_logits_ptr = qk_max;
+    float* exp_sums_ptr = exp_sums + seq_idx * num_heads * max_num_partitions +
+                          head_idx * max_num_partitions + partition_idx;
+    *exp_sums_ptr = exp_sum;
+  }
+
+  // Each thread will fetch 16 bytes from the value cache at a time.
+  constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
+  using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
+  using L_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
+  using V_quant_vec = typename Vec<cache_t, V_VEC_SIZE>::Type;
+  using Float_L_vec = typename FloatVec<L_vec>::Type;
+
+  constexpr int NUM_V_VECS_PER_ROW = BLOCK_SIZE / V_VEC_SIZE;
+  constexpr int NUM_ROWS_PER_ITER = WARP_SIZE / NUM_V_VECS_PER_ROW;
+  constexpr int NUM_ROWS_PER_THREAD =
+      DIVIDE_ROUND_UP(HEAD_SIZE, NUM_ROWS_PER_ITER);
+
+  // NOTE(woosuk): We use FP32 for the accumulator for better accuracy.
+  float accs[NUM_ROWS_PER_THREAD];
+#pragma unroll
+  for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+    accs[i] = 0.f;
+  }
+
+  scalar_t zero_value;
+  zero(zero_value);
+  for (int block_idx = start_block_idx + warp_idx; block_idx < end_block_idx;
+       block_idx += NUM_WARPS) {
+    // NOTE(woosuk): The block number is stored in int32. However, we cast it to
+    // int64 because int32 can lead to overflow when this variable is multiplied
+    // by large numbers (e.g., kv_block_stride).
+    // For blocksparse attention: skip computation on blocks that are not
+    // attended
+    if constexpr (IS_BLOCK_SPARSE) {
+      int v_bs_block_id = block_idx * BLOCK_SIZE / blocksparse_block_size;
+      if (!((v_bs_block_id + bs_block_offset) % blocksparse_vert_stride == 0) &&
+          !((v_bs_block_id > q_bs_block_id - blocksparse_local_blocks))) {
+        continue;
+      }
+    }
+    const int64_t physical_block_number =
+        static_cast<int64_t>(block_table[block_idx]);
+    const int physical_block_offset = (lane % NUM_V_VECS_PER_ROW) * V_VEC_SIZE;
+    const int token_idx = block_idx * BLOCK_SIZE + physical_block_offset;
+    L_vec logits_vec;
+    from_float(logits_vec, *reinterpret_cast<Float_L_vec*>(logits + token_idx -
+                                                           start_token_idx));
+
+    const cache_t* v_ptr = v_cache + physical_block_number * kv_block_stride +
+                           kv_head_idx * kv_head_stride;
+#pragma unroll
+    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+      const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+      if (row_idx < HEAD_SIZE) {
+        const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
+        V_vec v_vec;
+
+        if constexpr (KV_DTYPE == Fp8KVCacheDataType::kAuto) {
+          v_vec = *reinterpret_cast<const V_vec*>(v_ptr + offset);
+        } else {
+          V_quant_vec v_quant_vec =
+              *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
+          // Vector conversion from V_quant_vec to V_vec.
+          v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
+                                                                    *v_scale);
+        }
+        if (block_idx == num_seq_blocks - 1) {
+          // NOTE(woosuk): When v_vec contains the tokens that are out of the
+          // context, we should explicitly zero out the values since they may
+          // contain NaNs. See
+          // https://github.com/vllm-project/vllm/issues/641#issuecomment-1682544472
+          scalar_t* v_vec_ptr = reinterpret_cast<scalar_t*>(&v_vec);
+#pragma unroll
+          for (int j = 0; j < V_VEC_SIZE; j++) {
+            v_vec_ptr[j] = token_idx + j < seq_len ? v_vec_ptr[j] : zero_value;
+          }
+        }
+        accs[i] += dot(logits_vec, v_vec);
+      }
+    }
+  }
+
+  // Perform reduction within each warp.
+#pragma unroll
+  for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+    float acc = accs[i];
+#pragma unroll
+    for (int mask = NUM_V_VECS_PER_ROW / 2; mask >= 1; mask /= 2) {
+      acc += VLLM_SHFL_XOR_SYNC(acc, mask);
+    }
+    accs[i] = acc;
+  }
+
+  // NOTE(woosuk): A barrier is required because the shared memory space for
+  // logits is reused for the output.
+  __syncthreads();
+
+  // Perform reduction across warps.
+  float* out_smem = reinterpret_cast<float*>(shared_mem);
+#pragma unroll
+  for (int i = NUM_WARPS; i > 1; i /= 2) {
+    int mid = i / 2;
+    // Upper warps write to shared memory.
+    if (warp_idx >= mid && warp_idx < i) {
+      float* dst = &out_smem[(warp_idx - mid) * HEAD_SIZE];
+#pragma unroll
+      for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+        const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+        if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
+          dst[row_idx] = accs[i];
+        }
+      }
+    }
+    __syncthreads();
+
+    // Lower warps update the output.
+    if (warp_idx < mid) {
+      const float* src = &out_smem[warp_idx * HEAD_SIZE];
+#pragma unroll
+      for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+        const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+        if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
+          accs[i] += src[row_idx];
+        }
+      }
+    }
+    __syncthreads();
+  }
+
+  // Write the final output.
+  if (warp_idx == 0) {
+    scalar_t* out_ptr =
+        out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+        head_idx * max_num_partitions * HEAD_SIZE + partition_idx * HEAD_SIZE;
+#pragma unroll
+    for (int i = 0; i < NUM_ROWS_PER_THREAD; i++) {
+      const int row_idx = lane / NUM_V_VECS_PER_ROW + i * NUM_ROWS_PER_ITER;
+      if (row_idx < HEAD_SIZE && lane % NUM_V_VECS_PER_ROW == 0) {
+        from_float(*(out_ptr + row_idx), accs[i]);
+      }
+    }
+  }
+}
+
+// Grid: (num_heads, num_seqs, 1).
+template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
+          int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
+          bool IS_BLOCK_SPARSE>
+__global__ void paged_attention_v1_kernel(
+    scalar_t* __restrict__ out,           // [num_seqs, num_heads, head_size]
+    const scalar_t* __restrict__ q,       // [num_seqs, num_heads, head_size]
+    const cache_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                          // head_size/x, block_size, x]
+    const cache_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                          // head_size, block_size]
+    const int num_kv_heads,               // [num_heads]
+    const float scale,
+    const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_blocks_per_seq,
+    const float* __restrict__ alibi_slopes,  // [num_heads]
+    const int q_stride, const int kv_block_stride, const int kv_head_stride,
+    const float* k_scale, const float* v_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
+                         KV_DTYPE, IS_BLOCK_SPARSE>(
+      /* exp_sums */ nullptr, /* max_logits */ nullptr, out, q, k_cache,
+      v_cache, num_kv_heads, scale, block_tables, seq_lens,
+      max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride,
+      kv_head_stride, k_scale, v_scale, tp_rank, blocksparse_local_blocks,
+      blocksparse_vert_stride, blocksparse_block_size,
+      blocksparse_head_sliding_step);
+}
+
+// Grid: (num_heads, num_seqs, max_num_partitions).
+template <typename scalar_t, typename cache_t, int HEAD_SIZE, int BLOCK_SIZE,
+          int NUM_THREADS, vllm::Fp8KVCacheDataType KV_DTYPE,
+          bool IS_BLOCK_SPARSE,
+          int PARTITION_SIZE>
+__global__ void paged_attention_v2_kernel(
+    float* __restrict__ exp_sums,  // [num_seqs, num_heads, max_num_partitions]
+    float* __restrict__ max_logits,       // [num_seqs, num_heads,
+                                          // max_num_partitions]
+    scalar_t* __restrict__ tmp_out,       // [num_seqs, num_heads,
+                                          // max_num_partitions, head_size]
+    const scalar_t* __restrict__ q,       // [num_seqs, num_heads, head_size]
+    const cache_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                          // head_size/x, block_size, x]
+    const cache_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                          // head_size, block_size]
+    const int num_kv_heads,               // [num_heads]
+    const float scale,
+    const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_blocks_per_seq,
+    const float* __restrict__ alibi_slopes,  // [num_heads]
+    const int q_stride, const int kv_block_stride, const int kv_head_stride,
+    const float* k_scale, const float* v_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
+                         KV_DTYPE, IS_BLOCK_SPARSE, PARTITION_SIZE>(
+      exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
+      block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
+      kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
+      blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
+      blocksparse_head_sliding_step);
+}
+
+// Grid: (num_heads, num_seqs).
+template <typename scalar_t, int HEAD_SIZE, int NUM_THREADS,
+          int PARTITION_SIZE>
+__global__ void paged_attention_v2_reduce_kernel(
+    scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+    const float* __restrict__ exp_sums,    // [num_seqs, num_heads,
                                            // max_num_partitions]
-     scalar_t* __restrict__ tmp_out,       // [num_seqs, num_heads,
+    const float* __restrict__ max_logits,  // [num_seqs, num_heads,
+                                           // max_num_partitions]
+    const scalar_t* __restrict__ tmp_out,  // [num_seqs, num_heads,
                                            // max_num_partitions, head_size]
-     const scalar_t* __restrict__ q,       // [num_seqs, num_heads, head_size]
-     const cache_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
-                                           // head_size/x, block_size, x]
-     const cache_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
-                                           // head_size, block_size]
-     const int num_kv_heads,               // [num_heads]
-     const float scale,
-     const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
-     const int* __restrict__ seq_lens,      // [num_seqs]
-     const int max_num_blocks_per_seq,
-     const float* __restrict__ alibi_slopes,  // [num_heads]
-     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-     const float* k_scale, const float* v_scale, const int tp_rank,
-     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
-   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
-                          KV_DTYPE, IS_BLOCK_SPARSE, PARTITION_SIZE>(
-       exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
-       block_tables, seq_lens, max_num_blocks_per_seq, alibi_slopes, q_stride,
-       kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,
-       blocksparse_local_blocks, blocksparse_vert_stride, blocksparse_block_size,
-       blocksparse_head_sliding_step);
- }
- 
- // Grid: (num_heads, num_seqs).
- template <typename scalar_t, int HEAD_SIZE, int NUM_THREADS,
-           int PARTITION_SIZE>
- __global__ void paged_attention_v2_reduce_kernel(
-     scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
-     const float* __restrict__ exp_sums,    // [num_seqs, num_heads,
-                                            // max_num_partitions]
-     const float* __restrict__ max_logits,  // [num_seqs, num_heads,
-                                            // max_num_partitions]
-     const scalar_t* __restrict__ tmp_out,  // [num_seqs, num_heads,
-                                            // max_num_partitions, head_size]
-     const int* __restrict__ seq_lens,      // [num_seqs]
-     const int max_num_partitions) {
-   const int num_heads = gridDim.x;
-   const int head_idx = blockIdx.x;
-   const int seq_idx = blockIdx.y;
-   const int seq_len = seq_lens[seq_idx];
-   const int num_partitions = DIVIDE_ROUND_UP(seq_len, PARTITION_SIZE);
-   if (num_partitions == 1) {
-     // No need to reduce. Only copy tmp_out to out.
-     scalar_t* out_ptr =
-         out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
-     const scalar_t* tmp_out_ptr =
-         tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
-         head_idx * max_num_partitions * HEAD_SIZE;
-     for (int i = threadIdx.x; i < HEAD_SIZE; i += blockDim.x) {
-       out_ptr[i] = tmp_out_ptr[i];
-     }
-     // Terminate the thread block.
-     return;
-   }
- 
-   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-   const int warp_idx = threadIdx.x / WARP_SIZE;
-   const int lane = threadIdx.x % WARP_SIZE;
- 
-   // Size: 2 * num_partitions.
-   extern __shared__ char shared_mem[];
-   // Workspace for reduction.
-   __shared__ float red_smem[2 * NUM_WARPS];
- 
-   // Load max logits to shared memory.
-   float* shared_max_logits = reinterpret_cast<float*>(shared_mem);
-   const float* max_logits_ptr = max_logits +
-                                 seq_idx * num_heads * max_num_partitions +
-                                 head_idx * max_num_partitions;
-   float max_logit = -FLT_MAX;
-   for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
-     const float l = max_logits_ptr[i];
-     shared_max_logits[i] = l;
-     max_logit = fmaxf(max_logit, l);
-   }
-   __syncthreads();
- 
-   // Get the global max logit.
-   // Reduce within the warp.
- #pragma unroll
-   for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
-     max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
-   }
-   if (lane == 0) {
-     red_smem[warp_idx] = max_logit;
-   }
-   __syncthreads();
-   // Reduce across warps.
-   max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
- #pragma unroll
-   for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
-     max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
-   }
-   // Broadcast the max value to all threads.
-   max_logit = VLLM_SHFL_SYNC(max_logit, 0);
- 
-   // Load rescaled exp sums to shared memory.
-   float* shared_exp_sums =
-       reinterpret_cast<float*>(shared_mem + sizeof(float) * num_partitions);
-   const float* exp_sums_ptr = exp_sums +
-                               seq_idx * num_heads * max_num_partitions +
-                               head_idx * max_num_partitions;
-   float global_exp_sum = 0.0f;
-   for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
-     float l = shared_max_logits[i];
-     float rescaled_exp_sum = exp_sums_ptr[i] * expf(l - max_logit);
-     global_exp_sum += rescaled_exp_sum;
-     shared_exp_sums[i] = rescaled_exp_sum;
-   }
-   __syncthreads();
-   global_exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], global_exp_sum);
-   const float inv_global_exp_sum = __fdividef(1.0f, global_exp_sum + 1e-6f);
- 
-   // Aggregate tmp_out to out.
-   const scalar_t* tmp_out_ptr =
-       tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
-       head_idx * max_num_partitions * HEAD_SIZE;
-   scalar_t* out_ptr =
-       out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
- #pragma unroll
-   for (int i = threadIdx.x; i < HEAD_SIZE; i += NUM_THREADS) {
-     float acc = 0.0f;
-     for (int j = 0; j < num_partitions; ++j) {
-       acc += to_float(tmp_out_ptr[j * HEAD_SIZE + i]) * shared_exp_sums[j] *
-              inv_global_exp_sum;
-     }
-     from_float(out_ptr[i], acc);
-   }
- }
- 
- }  // namespace vllm
- 
- #undef WARP_SIZE
- #undef MAX
- #undef MIN
- #undef DIVIDE_ROUND_UP 
\ No newline at end of file
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_partitions) {
+  const int num_heads = gridDim.x;
+  const int head_idx = blockIdx.x;
+  const int seq_idx = blockIdx.y;
+  const int seq_len = seq_lens[seq_idx];
+  const int num_partitions = DIVIDE_ROUND_UP(seq_len, PARTITION_SIZE);
+  if (num_partitions == 1) {
+    // No need to reduce. Only copy tmp_out to out.
+    scalar_t* out_ptr =
+        out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+    const scalar_t* tmp_out_ptr =
+        tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+        head_idx * max_num_partitions * HEAD_SIZE;
+    for (int i = threadIdx.x; i < HEAD_SIZE; i += blockDim.x) {
+      out_ptr[i] = tmp_out_ptr[i];
+    }
+    // Terminate the thread block.
+    return;
+  }
+
+  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  const int warp_idx = threadIdx.x / WARP_SIZE;
+  const int lane = threadIdx.x % WARP_SIZE;
+
+  // Size: 2 * num_partitions.
+  extern __shared__ char shared_mem[];
+  // Workspace for reduction.
+  __shared__ float red_smem[2 * NUM_WARPS];
+
+  // Load max logits to shared memory.
+  float* shared_max_logits = reinterpret_cast<float*>(shared_mem);
+  const float* max_logits_ptr = max_logits +
+                                seq_idx * num_heads * max_num_partitions +
+                                head_idx * max_num_partitions;
+  float max_logit = -FLT_MAX;
+  for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
+    const float l = max_logits_ptr[i];
+    shared_max_logits[i] = l;
+    max_logit = fmaxf(max_logit, l);
+  }
+  __syncthreads();
+
+  // Get the global max logit.
+  // Reduce within the warp.
+#pragma unroll
+  for (int mask = WARP_SIZE / 2; mask >= 1; mask /= 2) {
+    max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
+  }
+  if (lane == 0) {
+    red_smem[warp_idx] = max_logit;
+  }
+  __syncthreads();
+  // Reduce across warps.
+  max_logit = lane < NUM_WARPS ? red_smem[lane] : -FLT_MAX;
+#pragma unroll
+  for (int mask = NUM_WARPS / 2; mask >= 1; mask /= 2) {
+    max_logit = fmaxf(max_logit, VLLM_SHFL_XOR_SYNC(max_logit, mask));
+  }
+  // Broadcast the max value to all threads.
+  max_logit = VLLM_SHFL_SYNC(max_logit, 0);
+
+  // Load rescaled exp sums to shared memory.
+  float* shared_exp_sums =
+      reinterpret_cast<float*>(shared_mem + sizeof(float) * num_partitions);
+  const float* exp_sums_ptr = exp_sums +
+                              seq_idx * num_heads * max_num_partitions +
+                              head_idx * max_num_partitions;
+  float global_exp_sum = 0.0f;
+  for (int i = threadIdx.x; i < num_partitions; i += blockDim.x) {
+    float l = shared_max_logits[i];
+    float rescaled_exp_sum = exp_sums_ptr[i] * expf(l - max_logit);
+    global_exp_sum += rescaled_exp_sum;
+    shared_exp_sums[i] = rescaled_exp_sum;
+  }
+  __syncthreads();
+  global_exp_sum = block_sum<NUM_WARPS>(&red_smem[NUM_WARPS], global_exp_sum);
+  const float inv_global_exp_sum = __fdividef(1.0f, global_exp_sum + 1e-6f);
+
+  // Aggregate tmp_out to out.
+  const scalar_t* tmp_out_ptr =
+      tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+      head_idx * max_num_partitions * HEAD_SIZE;
+  scalar_t* out_ptr =
+      out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+#pragma unroll
+  for (int i = threadIdx.x; i < HEAD_SIZE; i += NUM_THREADS) {
+    float acc = 0.0f;
+    for (int j = 0; j < num_partitions; ++j) {
+      acc += to_float(tmp_out_ptr[j * HEAD_SIZE + i]) * shared_exp_sums[j] *
+             inv_global_exp_sum;
+    }
+    from_float(out_ptr[i], acc);
+  }
+}
+
+}  // namespace vllm
+
+#undef WARP_SIZE
+#undef MAX
+#undef MIN
+#undef DIVIDE_ROUND_UP
\ No newline at end of file

csrc/attention/vertical_slash_index.cu

+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT license.
+
+#include <assert.h>
+
+#include <cuda.h>
+
+#include <torch/all.h>
+
+__device__ int64_t save_blocks(int* block_offset, int64_t range_start,
+                               int64_t range_end, int64_t block_size,
+                               int64_t input_block_count, int64_t kv_seqlen) {
+  if (range_start >= kv_seqlen) {
+    return input_block_count;
+  }
+  if (range_end > kv_seqlen) {
+    range_end = kv_seqlen;
+  }
+  int64_t current_block_count = input_block_count;
+  for (int idx = range_start; idx < range_end; idx += block_size) {
+    block_offset[current_block_count++] = idx;
+  }
+  return current_block_count;
+}
+
+__global__ void convert_vertical_slash_indexes_kernel(
+    const int* q_seqlens,         // [BATCH, ]
+    const int* kv_seqlens,        // [BATCH, ]
+    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
+    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
+    int* block_count,             // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
+    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
+    int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
+    int64_t NNZ_V, int64_t NNZ_S,
+    bool causal  // True for intra, False for succ
+) {
+  const int batch_idx = blockIdx.y;
+  const int head_idx = blockIdx.x;
+  const int group_idx = blockIdx.z;
+
+  int64_t q_seqlen = q_seqlens[batch_idx];
+  int64_t kv_seqlen = kv_seqlens[batch_idx];
+  int64_t block_idx_m = group_idx * blockDim.x + threadIdx.x;
+  int64_t start_m = block_idx_m * BLOCK_SIZE_M;
+  if (start_m >= q_seqlen) {
+    return;
+  }
+  int64_t end_m = start_m + BLOCK_SIZE_M;
+  vertical_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_V;
+  slash_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_S;
+  int64_t row_offset = (batch_idx * N_HEADS + head_idx) * N_ROWS + block_idx_m;
+  block_count += row_offset;
+  block_offset += row_offset * NNZ_S;
+  column_count += row_offset;
+  column_index += row_offset * NNZ_V;
+
+  bool has_slash = true;
+  int64_t tmp_col_cnt = 0, tmp_blk_cnt = 0;
+  int64_t s = 0, v = 0;
+  int64_t v_idx = vertical_indexes[v++];
+  int64_t s_idx = slash_indexes[s++];
+  if (causal) {
+    while (s_idx >= end_m + (kv_seqlen - q_seqlen) && s < NNZ_S) {
+      s_idx = slash_indexes[s++];
+    }
+    if (s_idx > end_m + (kv_seqlen - q_seqlen)) has_slash = false;
+    s_idx = max((kv_seqlen - q_seqlen) + end_m - s_idx, BLOCK_SIZE_M);
+  } else {
+    while (s_idx >= end_m + kv_seqlen && s < NNZ_S) {
+      s_idx = slash_indexes[s++];
+    }
+    if (s_idx > end_m + kv_seqlen) has_slash = false;
+    s_idx = max(kv_seqlen + end_m - s_idx, BLOCK_SIZE_M);
+  }
+
+  int64_t range_start = s_idx - BLOCK_SIZE_M, range_end = s_idx;
+  if (!has_slash) {
+    if (causal) {
+      range_start = (kv_seqlen - q_seqlen) + end_m;
+      range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
+    } else {
+      range_start = kv_seqlen;
+      range_end = kv_seqlen + BLOCK_SIZE_N;
+    }
+  }
+
+  bool slash_finished = false;
+  while (1) {
+    if (v_idx < range_end) {
+      if (v_idx < range_start) {
+        column_index[tmp_col_cnt++] = v_idx;
+      }
+      if (v < NNZ_V) {
+        v_idx = vertical_indexes[v++];
+      } else {
+        if (causal)
+          v_idx = end_m + BLOCK_SIZE_N + (kv_seqlen - q_seqlen);
+        else
+          v_idx = end_m + BLOCK_SIZE_N + kv_seqlen;
+      }
+    } else {
+      if ((s < NNZ_S && causal) ||
+          (s < NNZ_S && !causal && slash_indexes[s] >= start_m)) {
+        if (causal)
+          s_idx = max((kv_seqlen - q_seqlen) + end_m - slash_indexes[s++],
+                      BLOCK_SIZE_M);
+        else
+          s_idx = max(kv_seqlen + end_m - slash_indexes[s++], BLOCK_SIZE_M);
+      } else {
+        if (v == NNZ_V || (v_idx > range_start && causal)) {
+          // add the last vertical if no more slash
+          if (v == NNZ_V && !causal && v_idx < kv_seqlen) {
+            column_index[tmp_col_cnt++] = v_idx;
+          }
+          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
+                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
+          break;
+        } else {
+          if (causal) {
+            range_start = (kv_seqlen - q_seqlen) + end_m;
+            range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
+          } else {
+            // if slash_finished but there are vertical left, save current
+            // blocks
+            tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
+                                      BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
+            range_start = kv_seqlen;
+            range_end = kv_seqlen + BLOCK_SIZE_N;
+          }
+          slash_finished = true;
+        }
+      }
+      if (!slash_finished) {
+        if (s_idx > range_end + BLOCK_SIZE_M) {
+          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
+                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
+          range_start = s_idx - BLOCK_SIZE_M;
+          range_end = s_idx;
+        } else if (s_idx > range_end) {
+          range_end += BLOCK_SIZE_M;
+        }
+      }
+    }
+  }
+
+  block_count[0] = tmp_blk_cnt;
+  column_count[0] = tmp_col_cnt;
+}
+
+void convert_vertical_slash_indexes_64x64(
+    const int* q_seqlens,         // [BATCH, ]
+    const int* kv_seqlens,        // [BATCH, ]
+    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
+    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
+    int* block_count,             // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
+    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
+    int64_t BATCH_SIZE, int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M,
+    int64_t BLOCK_SIZE_N, int64_t NNZ_V, int64_t NNZ_S, bool causal) {
+  const int N_THREADS = 64;
+  const dim3 dimBlock(N_THREADS);
+  const dim3 dimGrid(N_HEADS, BATCH_SIZE, (N_ROWS + N_THREADS - 1) / N_THREADS);
+  convert_vertical_slash_indexes_kernel<<<dimGrid, dimBlock>>>(
+      q_seqlens, kv_seqlens, vertical_indexes, slash_indexes, block_count,
+      block_offset, column_count, column_index, N_HEADS, N_ROWS, BLOCK_SIZE_M,
+      BLOCK_SIZE_N, NNZ_V, NNZ_S, causal);
+}
+
+/**
+ * Implements the Algorithm 4 in paper https://arxiv.org/abs/2407.02490.
+ *
+ * This function builds the index of each row of blocks from vertical indices
+ * and slash indices. The vertical indices are treated as points, while the
+ * slash indices are converted as ranges. The output consists of the merged
+ * ranges and separate column indices, where the ranges are represented by
+ * block indices.
+ *
+ * The implementation is referenced from the original MInference repo:
+ * https://github.com/microsoft/MInference/blob/main/csrc/vertical_slash_index.cu.
+ */
+void convert_vertical_slash_indexes(
+    torch::Tensor& block_count,      // [BATCH, N_HEADS, NUM_ROWS]
+    torch::Tensor& block_offset,     // [BATCH, N_HEADS, NUM_ROWS, NNZ_S]
+    torch::Tensor& column_count,     // [BATCH, N_HEADS, NUM_ROWS]
+    torch::Tensor& column_index,     // [BATCH, N_HEADS, NUM_ROWS, NNZ_V]
+    torch::Tensor q_seqlens,         // [BATCH, ]
+    torch::Tensor kv_seqlens,        // [BATCH, ]
+    torch::Tensor vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
+    torch::Tensor slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
+    int64_t context_size, int64_t block_size_M, int64_t block_size_N,
+    bool causal) {
+  cudaSetDevice(q_seqlens.get_device());
+
+  int batch_size = slash_indexes.size(0);
+  int num_heads = slash_indexes.size(1);
+  int nnz_slash = slash_indexes.size(2);
+  int nnz_vertical = vertical_indexes.size(2);
+  int num_rows = (context_size + block_size_M - 1) / block_size_M;
+
+  convert_vertical_slash_indexes_64x64(
+      q_seqlens.data_ptr<int>(), kv_seqlens.data_ptr<int>(),
+      vertical_indexes.data_ptr<int>(), slash_indexes.data_ptr<int>(),
+      block_count.data_ptr<int>(), block_offset.data_ptr<int>(),
+      column_count.data_ptr<int>(), column_index.data_ptr<int>(), batch_size,
+      num_heads, num_rows, block_size_M, block_size_N, nnz_vertical, nnz_slash,
+      causal);
+}
+
+__global__ void convert_vertical_slash_indexes_kernel_mergehead(
+    const int* q_seqlens,         // [BATCH, ]
+    const int* kv_seqlens,        // [BATCH, ]
+    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
+    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
+    const int* per_head_vertical_topkv, const int* per_head_slash_topkv,
+    int* block_count,   // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
+    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
+    int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
+    int64_t NNZ_V, int64_t NNZ_S,
+    bool causal  // True for intra, False for succ
+) {
+  const int batch_idx = blockIdx.y;
+  const int head_idx = blockIdx.x;
+  const int group_idx = blockIdx.z;
+
+  int64_t q_seqlen = q_seqlens[batch_idx];
+  int64_t kv_seqlen = kv_seqlens[batch_idx];
+  int64_t block_idx_m = group_idx * blockDim.x + threadIdx.x;
+  int64_t start_m = block_idx_m * BLOCK_SIZE_M;
+  if (start_m >= q_seqlen) {
+    return;
+  }
+  int64_t end_m = start_m + BLOCK_SIZE_M;
+  vertical_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_V;
+  slash_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_S;
+  int64_t row_offset = (batch_idx * N_HEADS + head_idx) * N_ROWS + block_idx_m;
+  block_count += row_offset;
+  block_offset += row_offset * NNZ_S;
+  column_count += row_offset;
+  column_index += row_offset * NNZ_V;
+
+  // MergeHead: each head has it's unique max topk NNZ_V，NNZ_S. (NNZ_V，NNZ_S
+  // above is buffer size, use to compute offset)
+  NNZ_S = per_head_slash_topkv[head_idx];
+  NNZ_V = per_head_vertical_topkv[head_idx];
+
+  bool has_slash = true;
+  int64_t tmp_col_cnt = 0, tmp_blk_cnt = 0;
+  int64_t s = 0, v = 0;
+  int64_t v_idx = vertical_indexes[v++];
+  int64_t s_idx = slash_indexes[s++];
+  if (causal) {
+    while (s_idx >= end_m + (kv_seqlen - q_seqlen) && s < NNZ_S) {
+      s_idx = slash_indexes[s++];
+    }
+    if (s_idx > end_m + (kv_seqlen - q_seqlen)) has_slash = false;
+    s_idx = max((kv_seqlen - q_seqlen) + end_m - s_idx, BLOCK_SIZE_M);
+  } else {
+    while (s_idx >= end_m + kv_seqlen && s < NNZ_S) {
+      s_idx = slash_indexes[s++];
+    }
+    if (s_idx > end_m + kv_seqlen) has_slash = false;
+    s_idx = max(kv_seqlen + end_m - s_idx, BLOCK_SIZE_M);
+  }
+
+  int64_t range_start = s_idx - BLOCK_SIZE_M, range_end = s_idx;
+  if (!has_slash) {
+    if (causal) {
+      range_start = (kv_seqlen - q_seqlen) + end_m;
+      range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
+    } else {
+      range_start = kv_seqlen;
+      range_end = kv_seqlen + BLOCK_SIZE_N;
+    }
+  }
+
+  bool slash_finished = false;
+  while (1) {
+    if (v_idx < range_end) {
+      if (v_idx < range_start) {
+        column_index[tmp_col_cnt++] = v_idx;
+      }
+      if (v < NNZ_V) {
+        v_idx = vertical_indexes[v++];
+      } else {
+        if (causal)
+          v_idx = end_m + BLOCK_SIZE_N + (kv_seqlen - q_seqlen);
+        else
+          v_idx = end_m + BLOCK_SIZE_N + kv_seqlen;
+      }
+    } else {
+      if ((s < NNZ_S && causal) ||
+          (s < NNZ_S && !causal && slash_indexes[s] >= start_m)) {
+        if (causal)
+          s_idx = max((kv_seqlen - q_seqlen) + end_m - slash_indexes[s++],
+                      BLOCK_SIZE_M);
+        else
+          s_idx = max(kv_seqlen + end_m - slash_indexes[s++], BLOCK_SIZE_M);
+      } else {
+        if (v == NNZ_V || (v_idx > range_start && causal)) {
+          // add the last vertical if no more slash
+          if (v == NNZ_V && !causal && v_idx < kv_seqlen) {
+            column_index[tmp_col_cnt++] = v_idx;
+          }
+          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
+                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
+          break;
+        } else {
+          if (causal) {
+            range_start = (kv_seqlen - q_seqlen) + end_m;
+            range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
+          } else {
+            // if slash_finished but there are vertical left, save current
+            // blocks
+            tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
+                                      BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
+            range_start = kv_seqlen;
+            range_end = kv_seqlen + BLOCK_SIZE_N;
+          }
+          slash_finished = true;
+        }
+      }
+      if (!slash_finished) {
+        if (s_idx > range_end + BLOCK_SIZE_M) {
+          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
+                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
+          range_start = s_idx - BLOCK_SIZE_M;
+          range_end = s_idx;
+        } else if (s_idx > range_end) {
+          range_end += BLOCK_SIZE_M;
+        }
+      }
+    }
+  }
+
+  block_count[0] = tmp_blk_cnt;
+  column_count[0] = tmp_col_cnt;
+}
+
+void convert_vertical_slash_indexes_64x64_mergehead(
+    const int* q_seqlens,         // [BATCH, ]
+    const int* kv_seqlens,        // [BATCH, ]
+    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
+    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
+    int* per_head_vertical_topkv, int* per_head_slash_topkv,
+    int* block_count,   // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
+    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
+    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
+    int64_t BATCH_SIZE, int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M,
+    int64_t BLOCK_SIZE_N, int64_t NNZ_V, int64_t NNZ_S, bool causal) {
+  const int N_THREADS = 64;
+  const dim3 dimBlock(N_THREADS);
+  const dim3 dimGrid(N_HEADS, BATCH_SIZE, (N_ROWS + N_THREADS - 1) / N_THREADS);
+  convert_vertical_slash_indexes_kernel_mergehead<<<dimGrid, dimBlock>>>(
+      q_seqlens, kv_seqlens, vertical_indexes, slash_indexes,
+      per_head_vertical_topkv, per_head_slash_topkv, block_count, block_offset,
+      column_count, column_index, N_HEADS, N_ROWS, BLOCK_SIZE_M, BLOCK_SIZE_N,
+      NNZ_V, NNZ_S, causal);
+}
+
+/**
+ * Implements the Algorithm 4 in paper https://arxiv.org/abs/2407.02490.
+ *
+ * Like the above convert_vertical_slash_indexes, but with
+ * pre-computed vertical and slash counts.
+ */
+void convert_vertical_slash_indexes_mergehead(
+    torch::Tensor& block_count,            // [BATCH, N_HEADS, NUM_ROWS]
+    torch::Tensor& block_offset,           // [BATCH, N_HEADS, NUM_ROWS, NNZ_S]
+    torch::Tensor& column_count,           // [BATCH, N_HEADS, NUM_ROWS]
+    torch::Tensor& column_index,           // [BATCH, N_HEADS, NUM_ROWS, NNZ_V]
+    torch::Tensor q_seqlens,               // [BATCH, ]
+    torch::Tensor kv_seqlens,              // [BATCH, ]
+    torch::Tensor vertical_indexes,        // [BATCH, N_HEADS, NNZ_V]
+    torch::Tensor slash_indexes,           // [BATCH, N_HEADS, NNZ_S]
+    torch::Tensor vertical_indices_count,  // [N_HEADS, ]
+    torch::Tensor slash_indices_count,     // [N_HEADS, ]
+    int64_t context_size, int64_t block_size_M, int64_t block_size_N,
+    bool causal) {
+  cudaSetDevice(q_seqlens.get_device());
+
+  int batch_size = slash_indexes.size(0);
+  int num_heads = slash_indexes.size(1);
+  int nnz_slash = slash_indexes.size(2);
+  int nnz_vertical = vertical_indexes.size(2);
+  int num_rows = (context_size + block_size_M - 1) / block_size_M;
+
+  convert_vertical_slash_indexes_64x64_mergehead(
+      q_seqlens.data_ptr<int>(), kv_seqlens.data_ptr<int>(),
+      vertical_indexes.data_ptr<int>(), slash_indexes.data_ptr<int>(),
+      vertical_indices_count.data_ptr<int>(),
+      slash_indices_count.data_ptr<int>(), block_count.data_ptr<int>(),
+      block_offset.data_ptr<int>(), column_count.data_ptr<int>(),
+      column_index.data_ptr<int>(), batch_size, num_heads, num_rows,
+      block_size_M, block_size_N, nnz_vertical, nnz_slash, causal);
+}

csrc/core/math.hpp

@@ -7,3 +7,22 @@ inline constexpr uint32_t next_pow_2(uint32_t const num) {
   if (num <= 1) return num;
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
+
+template <typename A, typename B>
+static inline constexpr auto div_ceil(A a, B b) {
+  return (a + b - 1) / b;
+}
+
+// Round a down to the next multiple of b. The caller is responsible for making
+// sure that b is non-zero
+template <typename T>
+inline constexpr T round_to_previous_multiple_of(T a, T b) {
+  return a % b == 0 ? a : (a / b) * b;
+}
+
+// Round a up to the next multiple of b. The caller is responsible for making
+// sure that b is non-zero
+template <typename T>
+inline constexpr T round_to_next_multiple_of(T a, T b) {
+  return a % b == 0 ? a : ((a / b) + 1) * b;
+}

csrc/core/scalar_type.hpp

@@ -315,6 +315,8 @@ static inline constexpr auto kS8 = ScalarType::int_(8);
 static inline constexpr auto kU8 = ScalarType::uint(8);
 static inline constexpr auto kU8B128 = ScalarType::uint(8, 128);
 
+static inline constexpr auto kFE2M1f =
+    ScalarType::float_(2, 1, true, ScalarType::NAN_NONE);
 static inline constexpr auto kFE3M2f =
     ScalarType::float_(3, 2, true, ScalarType::NAN_NONE);
 static inline constexpr auto kFE4M3fn =
@@ -332,6 +334,7 @@ static inline constexpr auto kInt8 = kS8;
 static inline constexpr auto kUint8 = kU8;
 static inline constexpr auto kUint8b128 = kU8B128;
 
+static inline constexpr auto kFloat4_e2m1f = kFE2M1f;
 static inline constexpr auto kFloat6_e3m2f = kFE3M2f;
 static inline constexpr auto kFloat8_e4m3fn = kFE4M3fn;
 static inline constexpr auto kFloat8_e5m2 = kFE5M2;

csrc/cpu/cpu_types_vsx.hpp

@@ -4,6 +4,7 @@
 
 #include <altivec.h>
 #include <cmath>
+#include <algorithm>
 #include <torch/all.h>
 
 namespace vec_op {
@@ -62,6 +63,10 @@ typedef struct f32x4x4_t {
   __vector float val[4];
 } f32x4x4_t;
 
+typedef struct i32x4x4_t {
+  __vector int32_t val[4];
+} i32x4x4_t;
+
 struct FP32Vec8;
 struct FP32Vec16;
 
@@ -98,6 +103,28 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     vec_xst(reg.val[0], 0, (signed short*)ptr);
     vec_xst(reg.val[1], 16, (signed short*)ptr);
   }
+
+  void save(void* ptr, const int elem_num) const {
+    const int clamped_elem = std::max(0, std::min(elem_num, 16));
+
+    // Calculate elements to store in each 128-bit part (8 elements each)
+    const int elements_val0 = std::min(clamped_elem, 8);
+    const int elements_val1 = std::max(clamped_elem - 8, 0);
+
+    // Convert elements to bytes (2 bytes per element)
+    const size_t bytes_val0 = elements_val0 * sizeof(signed short);
+    const size_t bytes_val1 = elements_val1 * sizeof(signed short);
+
+    signed short* dest = static_cast<signed short*>(ptr);
+    // Store the first part using vec_xst_len
+    if (bytes_val0 > 0) {
+      vec_xst_len(reg.val[0], dest, bytes_val0);
+    }
+    // Store the second part if needed
+    if (bytes_val1 > 0) {
+      vec_xst_len(reg.val[1], dest + elements_val0, bytes_val1);
+    }
+  }
 };
 
 const static __vector signed short zero = vec_splats((signed short)0);
@@ -257,6 +284,64 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   }
 };
 
+struct INT32Vec16 : public Vec<INT32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    i32x4x4_t reg;
+    int32_t values[VEC_ELEM_NUM];
+  };
+
+  i32x4x4_t reg;
+
+  explicit INT32Vec16(const void* data_ptr) {
+    reg.val[0] = vec_xl(0, reinterpret_cast<const __vector int32_t*>(data_ptr));
+    reg.val[1] =
+        vec_xl(16, reinterpret_cast<const __vector int32_t*>(data_ptr));
+    reg.val[2] =
+        vec_xl(32, reinterpret_cast<const __vector int32_t*>(data_ptr));
+    reg.val[3] =
+        vec_xl(48, reinterpret_cast<const __vector int32_t*>(data_ptr));
+  }
+
+  void save(int32_t* ptr) const {
+    vec_xst(reg.val[0], 0, reinterpret_cast<__vector int32_t*>(ptr));
+    vec_xst(reg.val[1], 16, reinterpret_cast<__vector int32_t*>(ptr));
+    vec_xst(reg.val[2], 32, reinterpret_cast<__vector int32_t*>(ptr));
+    vec_xst(reg.val[3], 48, reinterpret_cast<__vector int32_t*>(ptr));
+  }
+
+  void save(int32_t* ptr, const int elem_num) const {
+    const int elements_in_chunk1 =
+        (elem_num >= 0) ? ((elem_num >= 4) ? 4 : elem_num) : 0;
+    const int elements_in_chunk2 =
+        (elem_num > 4) ? ((elem_num >= 8) ? 4 : elem_num - 4) : 0;
+    const int elements_in_chunk3 =
+        (elem_num > 8) ? ((elem_num >= 12) ? 4 : elem_num - 8) : 0;
+    const int elements_in_chunk4 =
+        (elem_num > 12) ? ((elem_num >= 16) ? 4 : elem_num - 12) : 0;
+
+    const size_t bytes_chunk1 =
+        static_cast<size_t>(elements_in_chunk1 * sizeof(int32_t));
+    const size_t bytes_chunk2 =
+        static_cast<size_t>(elements_in_chunk2 * sizeof(int32_t));
+    const size_t bytes_chunk3 =
+        static_cast<size_t>(elements_in_chunk3 * sizeof(int32_t));
+    const size_t bytes_chunk4 =
+        static_cast<size_t>(elements_in_chunk4 * sizeof(int32_t));
+
+    vec_xst_len(reg.val[0], reinterpret_cast<int32_t*>(ptr), bytes_chunk1);
+    vec_xst_len(reg.val[1],
+                reinterpret_cast<int32_t*>(reinterpret_cast<char*>(ptr) + 16),
+                bytes_chunk2);
+    vec_xst_len(reg.val[2],
+                reinterpret_cast<int32_t*>(reinterpret_cast<char*>(ptr) + 32),
+                bytes_chunk3);
+    vec_xst_len(reg.val[3],
+                reinterpret_cast<int32_t*>(reinterpret_cast<char*>(ptr) + 48),
+                bytes_chunk4);
+  }
+};
+
 struct FP32Vec16 : public Vec<FP32Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
   union AliasReg {
@@ -319,6 +404,13 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
+  explicit FP32Vec16(const INT32Vec16& v) {
+    reg.val[0] = vec_ctf(v.reg.val[0], 0);
+    reg.val[1] = vec_ctf(v.reg.val[1], 0);
+    reg.val[2] = vec_ctf(v.reg.val[2], 0);
+    reg.val[3] = vec_ctf(v.reg.val[3], 0);
+  }
+
   FP32Vec16 operator*(const FP32Vec16& b) const {
     return FP32Vec16(f32x4x4_t({vec_mul(reg.val[0], b.reg.val[0]),
                                 vec_mul(reg.val[1], b.reg.val[1]),
@@ -347,6 +439,117 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                                 vec_div(reg.val[3], b.reg.val[3])}));
   }
 
+  FP32Vec16 clamp(const FP32Vec16& min, const FP32Vec16& max) const {
+    return FP32Vec16(f32x4x4_t(
+        {vec_min(max.reg.val[0], vec_max(min.reg.val[0], reg.val[0])),
+         vec_min(max.reg.val[1], vec_max(min.reg.val[1], reg.val[1])),
+         vec_min(max.reg.val[2], vec_max(min.reg.val[2], reg.val[2])),
+         vec_min(max.reg.val[3], vec_max(min.reg.val[3], reg.val[3]))}));
+  }
+
+  FP32Vec16 max(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_max(reg.val[0], b.reg.val[0]),
+                                vec_max(reg.val[1], b.reg.val[1]),
+                                vec_max(reg.val[2], b.reg.val[2]),
+                                vec_max(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 max(const FP32Vec16& b, int elem_num) const {
+    FP32Vec16 result;
+
+    // Create a vector of element indices for each chunk
+    __vector unsigned int indices = {0, 1, 2, 3};
+    __vector unsigned int elem_num_vec =
+        vec_splats(static_cast<unsigned int>(elem_num));
+
+    // Compute masks for each chunk
+    __vector unsigned int chunk_offset0 = {0, 0, 0,
+                                           0};  // Chunk 0: Elements 0-3
+    __vector unsigned int chunk_offset1 = {4, 4, 4,
+                                           4};  // Chunk 1: Elements 4-7
+    __vector unsigned int chunk_offset2 = {8, 8, 8,
+                                           8};  // Chunk 2: Elements 8-11
+    __vector unsigned int chunk_offset3 = {12, 12, 12,
+                                           12};  // Chunk 3: Elements 12-15
+
+    // Compute masks for each chunk
+    __vector bool int mask0 = vec_cmplt(indices + chunk_offset0, elem_num_vec);
+    __vector bool int mask1 = vec_cmplt(indices + chunk_offset1, elem_num_vec);
+    __vector bool int mask2 = vec_cmplt(indices + chunk_offset2, elem_num_vec);
+    __vector bool int mask3 = vec_cmplt(indices + chunk_offset3, elem_num_vec);
+
+    // Apply masks to compute the result for each chunk
+    result.reg.val[0] = vec_sel(this->reg.val[0],
+                                vec_max(this->reg.val[0], b.reg.val[0]), mask0);
+    result.reg.val[1] = vec_sel(this->reg.val[1],
+                                vec_max(this->reg.val[1], b.reg.val[1]), mask1);
+    result.reg.val[2] = vec_sel(this->reg.val[2],
+                                vec_max(this->reg.val[2], b.reg.val[2]), mask2);
+    result.reg.val[3] = vec_sel(this->reg.val[3],
+                                vec_max(this->reg.val[3], b.reg.val[3]), mask3);
+
+    return FP32Vec16(result.reg);
+  }
+
+  FP32Vec16 min(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_min(reg.val[0], b.reg.val[0]),
+                                vec_min(reg.val[1], b.reg.val[1]),
+                                vec_min(reg.val[2], b.reg.val[2]),
+                                vec_min(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 min(const FP32Vec16& b, int elem_num) const {
+    FP32Vec16 result;
+
+    vector unsigned int indices = {0, 1, 2, 3};
+    vector unsigned int elem_num_vec =
+        vec_splats(static_cast<unsigned int>(elem_num));
+
+    vector unsigned int chunk_offset0 = {0, 0, 0, 0};
+    vector unsigned int chunk_offset1 = {4, 4, 4, 4};
+    vector unsigned int chunk_offset2 = {8, 8, 8, 8};
+    vector unsigned int chunk_offset3 = {12, 12, 12, 12};
+
+    vector bool int mask0 = vec_cmplt(indices + chunk_offset0, elem_num_vec);
+    vector bool int mask1 = vec_cmplt(indices + chunk_offset1, elem_num_vec);
+    vector bool int mask2 = vec_cmplt(indices + chunk_offset2, elem_num_vec);
+    vector bool int mask3 = vec_cmplt(indices + chunk_offset3, elem_num_vec);
+
+    result.reg.val[0] = vec_sel(this->reg.val[0],
+                                vec_min(this->reg.val[0], b.reg.val[0]), mask0);
+    result.reg.val[1] = vec_sel(this->reg.val[1],
+                                vec_min(this->reg.val[1], b.reg.val[1]), mask1);
+    result.reg.val[2] = vec_sel(this->reg.val[2],
+                                vec_min(this->reg.val[2], b.reg.val[2]), mask2);
+    result.reg.val[3] = vec_sel(this->reg.val[3],
+                                vec_min(this->reg.val[3], b.reg.val[3]), mask3);
+
+    return FP32Vec16(result.reg);
+  }
+
+  FP32Vec16 abs() const {
+    return FP32Vec16(f32x4x4_t({vec_abs(reg.val[0]), vec_abs(reg.val[1]),
+                                vec_abs(reg.val[2]), vec_abs(reg.val[3])}));
+  }
+
+  float reduce_max() {
+    __vector float max01 = vec_max(reg.val[0], reg.val[1]);
+    __vector float max23 = vec_max(reg.val[2], reg.val[3]);
+    __vector float max_all = vec_max(max01, max23);
+    __vector float temp = vec_max(max_all, vec_sld(max_all, max_all, 8));
+    temp = vec_max(temp, vec_sld(temp, temp, 4));
+    return vec_extract(temp, 0);
+  }
+
+  float reduce_min() {
+    __vector float min01 = vec_min(reg.val[0], reg.val[1]);
+    __vector float min23 = vec_min(reg.val[2], reg.val[3]);
+    __vector float min_all = vec_min(min01, min23);
+    __vector float temp = vec_min(min_all, vec_sld(min_all, min_all, 8));
+    temp = vec_min(temp, vec_sld(temp, temp, 4));
+    return vec_extract(temp, 0);
+  }
+
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
@@ -377,6 +580,68 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     vec_xst(reg.val[2], 32, ptr);
     vec_xst(reg.val[3], 48, ptr);
   }
+
+  void save(float* ptr, const int elem_num) const {
+    const int elements_in_chunk1 =
+        (elem_num >= 0) ? ((elem_num >= 4) ? 4 : elem_num) : 0;
+    const int elements_in_chunk2 =
+        (elem_num > 4) ? ((elem_num >= 8) ? 4 : elem_num - 4) : 0;
+    const int elements_in_chunk3 =
+        (elem_num > 8) ? ((elem_num >= 12) ? 4 : elem_num - 8) : 0;
+    const int elements_in_chunk4 =
+        (elem_num > 12) ? ((elem_num >= 16) ? 4 : elem_num - 12) : 0;
+
+    const size_t bytes_chunk1 =
+        static_cast<size_t>(elements_in_chunk1 * sizeof(float));
+    const size_t bytes_chunk2 =
+        static_cast<size_t>(elements_in_chunk2 * sizeof(float));
+    const size_t bytes_chunk3 =
+        static_cast<size_t>(elements_in_chunk3 * sizeof(float));
+    const size_t bytes_chunk4 =
+        static_cast<size_t>(elements_in_chunk4 * sizeof(float));
+
+    vec_xst_len(reg.val[0], ptr, bytes_chunk1);
+    vec_xst_len(reg.val[1],
+                reinterpret_cast<float*>(reinterpret_cast<char*>(ptr) + 16),
+                bytes_chunk2);
+    vec_xst_len(reg.val[2],
+                reinterpret_cast<float*>(reinterpret_cast<char*>(ptr) + 32),
+                bytes_chunk3);
+    vec_xst_len(reg.val[3],
+                reinterpret_cast<float*>(reinterpret_cast<char*>(ptr) + 48),
+                bytes_chunk4);
+  }
+};
+
+struct INT8Vec16 : public Vec<INT8Vec16> {
+  constexpr static int VEC_NUM_ELEM = 16;  // 128 bits / 8 bits = 16
+
+  union AliasReg {
+    __vector signed char reg;
+    int8_t values[VEC_NUM_ELEM];
+  };
+
+  __vector signed char reg;
+
+  explicit INT8Vec16(const FP32Vec16& vec) {
+    __vector signed int ret[4];
+    ret[0] = vec_cts(vec.reg.val[0], 0);
+    ret[1] = vec_cts(vec.reg.val[1], 0);
+    ret[2] = vec_cts(vec.reg.val[2], 0);
+    ret[3] = vec_cts(vec.reg.val[3], 0);
+
+    __vector signed short packed1 = vec_packs(ret[0], ret[1]);
+    __vector signed short packed2 = vec_packs(ret[2], ret[3]);
+
+    reg = vec_packs(packed1, packed2);
+  }
+
+  void save(void* ptr) const {
+    *reinterpret_cast<__vector signed char*>(ptr) = reg;
+  }
+  void save(signed char* ptr, const int elem_num) {
+    vec_xst_len(reg, ptr, static_cast<size_t>(elem_num));
+  }
 };
 
 template <typename T>

csrc/cpu/pos_encoding.cpp

@@ -9,7 +9,8 @@ void rotary_embedding_impl(
     scalar_t* __restrict__ query,           /// [batch_size, seq_len, num_heads,
                                    /// head_size] or [num_tokens, num_heads,
                                    /// head_size]
-    scalar_t* __restrict__ key,  // [batch_size, seq_len, num_kv_heads,
+    scalar_t* __restrict__ key,  // nullptr (optional) or
+                                 // [batch_size, seq_len, num_kv_heads,
                                  // head_size] or [num_tokens, num_kv_heads,
                                  // head_size]
     const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
@@ -85,10 +86,13 @@ void rotary_embedding_impl(
       compute_loop(token_head, cache_ptr, query);
     }
 
-    for (int i = 0; i < num_kv_heads; ++i) {
-      const int head_idx = i;
-      const int64_t token_head = token_idx * key_stride + head_idx * head_size;
-      compute_loop(token_head, cache_ptr, key);
+    if (key != nullptr) {
+      for (int i = 0; i < num_kv_heads; ++i) {
+        const int head_idx = i;
+        const int64_t token_head =
+            token_idx * key_stride + head_idx * head_size;
+        compute_loop(token_head, cache_ptr, key);
+      }
     }
   }
 }
@@ -100,7 +104,8 @@ void rotary_embedding_gptj_impl(
     scalar_t* __restrict__ query,           /// [batch_size, seq_len, num_heads,
                                    /// head_size] or [num_tokens, num_heads,
                                    /// head_size]
-    scalar_t* __restrict__ key,  // [batch_size, seq_len, num_kv_heads,
+    scalar_t* __restrict__ key,  // nullptr (optional) or
+                                 // [batch_size, seq_len, num_kv_heads,
                                  // head_size] or [num_tokens, num_kv_heads,
                                  // head_size]
     const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
@@ -138,6 +143,10 @@ void rotary_embedding_gptj_impl(
     }
   }
 
+  if (key == nullptr) {
+    return;
+  }
+
 #pragma omp parallel for collapse(2)
   for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
     for (int i = 0; i < num_kv_heads; ++i) {
@@ -168,13 +177,13 @@ void rotary_embedding_gptj_impl(
 };  // namespace
 
 void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
-                      torch::Tensor& key, int64_t head_size,
+                      std::optional<torch::Tensor> key, int64_t head_size,
                       torch::Tensor& cos_sin_cache, bool is_neox) {
   int num_tokens = positions.numel();
   int rot_dim = cos_sin_cache.size(1);
   int num_heads = query.size(-1) / head_size;
-  int num_kv_heads = key.size(-1) / head_size;
-  int64_t key_stride = key.stride(-2);
+  int num_kv_heads = key.has_value() ? key->size(-1) / head_size : num_heads;
+  int64_t key_stride = key.has_value() ? key->stride(-2) : 0;
   int64_t query_stride = query.stride(-2);
 
   VLLM_DISPATCH_FLOATING_TYPES(
@@ -183,15 +192,15 @@ void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
         if (is_neox) {
           rotary_embedding_impl(
               positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
-              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
-              rot_dim, query_stride, key_stride, num_heads, num_kv_heads,
-              head_size, num_tokens);
+              key.has_value() ? key->data_ptr<scalar_t>() : nullptr,
+              cos_sin_cache.data_ptr<scalar_t>(), rot_dim, query_stride,
+              key_stride, num_heads, num_kv_heads, head_size, num_tokens);
         } else {
           rotary_embedding_gptj_impl(
               positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
-              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
-              rot_dim, query_stride, key_stride, num_heads, num_kv_heads,
-              head_size, num_tokens);
+              key.has_value() ? key->data_ptr<scalar_t>() : nullptr,
+              cos_sin_cache.data_ptr<scalar_t>(), rot_dim, query_stride,
+              key_stride, num_heads, num_kv_heads, head_size, num_tokens);
         }
 
         CPU_KERNEL_GUARD_OUT(rotary_embedding_impl)