[Triton Kernel] Add varlen segment mean triton kernel (#10369)

* upload triton kernel

* update kernel from review

* fix docstring

* update __init__

Author: l1cacheDell

paddlenlp/ops/triton_ops/__init__.py

@@ -17,6 +17,7 @@
         per_token_group_quant_fp8_api,
         per_token_group_quant_fp8_api_masked,
     )
+    from .segment_mean import segment_mean
 
     __all__ = +["per_token_group_quant_fp8_api_masked", "per_token_group_quant_fp8_api"]
 except:

paddlenlp/ops/triton_ops/segment_mean.py

@@ -0,0 +1,205 @@
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Notice: this kernel is especially implemented for the segment mean operation for a varlen qkv tensor.
+# for example, the k tensor is: [total_seqlen, num_head, head_dim],
+# where total_seqlen = seqlen 1 + seqlen 2 + ... + seqlen n.
+# So the segment mean triton kernel will do mean operation along the **seqlen** dim.
+# It will finally generate a `[bsz, num_head, head_dim]` shape-like result,
+# as the result of mean value of each seqlen segment.
+
+import paddle
+import triton.language as tl
+from paddle import _C_ops
+from paddle.base.framework import OpProtoHolder
+from paddle.base.layer_helper import LayerHelper
+from paddle.framework import in_dynamic_or_pir_mode
+
+from paddlenlp.ops.triton_ops.triton_utils import (
+    get_dtype_str,
+    paddle_use_triton,
+    rendering_common_template,
+)
+
+
+@paddle_use_triton(key=["num_heads", "head_dim"])
+def segmented_mean_reduce_kernel(
+    input_ptr,
+    output_ptr,
+    cu_seqlen_ptr,
+    num_batches,
+    num_heads,
+    head_dim,
+    input_stride_seq,
+    input_stride_head,
+    output_stride_batch,
+    output_stride_head,
+    BLOCK_SIZE_SEQ: tl.constexpr,
+    BLOCK_SIZE_HEAD: tl.constexpr,
+    BLOCK_SIZE_DIM: tl.constexpr,
+):
+    batch_idx = tl.program_id(0)
+    head_offset = tl.program_id(1) * BLOCK_SIZE_HEAD
+    dim_offset = tl.program_id(2) * BLOCK_SIZE_DIM
+
+    # 获取当前 segment 的 range
+    seq_start = tl.load(cu_seqlen_ptr + batch_idx)
+    seq_end = tl.load(cu_seqlen_ptr + batch_idx + 1)
+    seq_len = seq_end - seq_start
+
+    # head 和 dim 的实际索引（block中相对位置）
+    head_idx = head_offset + tl.arange(0, BLOCK_SIZE_HEAD)
+    dim_idx = dim_offset + tl.arange(0, BLOCK_SIZE_DIM)
+    mask_head = head_idx < num_heads
+    mask_dim = dim_idx < head_dim
+
+    # 初始化累加器（float32 精度）
+    acc = tl.zeros((BLOCK_SIZE_HEAD, BLOCK_SIZE_DIM), dtype=tl.float32)
+
+    for seq_offset in range(0, seq_len, BLOCK_SIZE_SEQ):
+        local_seq_idx = tl.arange(0, BLOCK_SIZE_SEQ)
+        mask_seq = local_seq_idx < (seq_len - seq_offset)
+        global_seq = seq_start + seq_offset + local_seq_idx
+        # shape: [BLOCK_SIZE_SEQ, BLOCK_SIZE_HEAD, BLOCK_SIZE_DIM]
+        input_ptrs = (
+            input_ptr
+            + global_seq[:, None, None] * input_stride_seq
+            + head_idx[None, :, None] * input_stride_head
+            + dim_idx[None, None, :]
+        )
+
+        # 加载输入，注意输入 dtype 指明 float16 以避免不必要转换
+        x = tl.load(
+            input_ptrs, mask=mask_seq[:, None, None] & mask_head[None, :, None] & mask_dim[None, None, :], other=0.0
+        ).to(tl.float32)
+
+        acc += tl.sum(x, axis=0)  # reduce over seq axis
+
+    mean = acc / seq_len
+
+    # 构造输出地址
+    output_ptrs = (
+        output_ptr + batch_idx * output_stride_batch + head_idx[:, None] * output_stride_head + dim_idx[None, :]
+    )
+
+    tl.store(output_ptrs, mean.to(input_ptr.dtype.element_ty), mask=mask_head[:, None] & mask_dim[None, :])
+
+
+def segment_mean(
+    x: paddle.Tensor, cu_seqlen: paddle.Tensor  # [total_seqlen, num_heads, head_dim]  # [batch_size + 1]
+):
+    """
+    Examples:
+        import paddle
+        from paddlenlp.ops.triton_ops.segment_mean import segment_mean
+
+        cu_seqlens = [0, 1024, 2048, 4096]
+        total_seqlen = 4096
+        num_head = 24
+        head_dim = 128
+        k = paddle.randn([total_seqlen, num_head, head_dim], dtype="float16")
+        cu_seqlen = paddle.to_tensor(cu_seqlens, paddle.int32)
+        km = segment_mean(k, cu_seqlen)
+    """
+    num_batches = cu_seqlen.shape[0] - 1
+
+    num_heads = x.shape[1]
+    head_dim = x.shape[2]
+
+    # 计算必要的strides
+    input_stride_seq = num_heads * head_dim
+    input_stride_head = head_dim
+
+    output_stride_batch = num_heads * head_dim
+    output_stride_head = head_dim
+    prepare_attr_for_triton_kernel = """
+    const int num_batches = cu_seqlen.shape()[0] - 1;
+    const int num_heads = x.shape()[1];
+    const int head_dim = x.shape()[2];
+    int input_stride_seq = num_heads * head_dim;
+    int input_stride_head = head_dim;
+    int output_stride_batch = num_heads * head_dim;
+    int output_stride_head = head_dim;
+    paddle::Tensor output_tensor = paddle::empty({num_batches, num_heads, head_dim}, x.dtype(), x.place());
+"""
+    op_name = "triton_segment_mean"
+    op_name += get_dtype_str(x.dtype)
+
+    # auto-tuning
+    segment_mean_configs = []
+    segment_mean_configs.append(
+        {"BLOCK_SIZE_SEQ": 128, "BLOCK_SIZE_HEAD": 4, "BLOCK_SIZE_DIM": 64, "num_stages": 2, "num_warps": 4}
+    )
+    segment_mean_configs.append(
+        {"BLOCK_SIZE_SEQ": 256, "BLOCK_SIZE_HEAD": 4, "BLOCK_SIZE_DIM": 64, "num_stages": 2, "num_warps": 4}
+    )
+    segment_mean_configs.append(
+        {"BLOCK_SIZE_SEQ": 512, "BLOCK_SIZE_HEAD": 8, "BLOCK_SIZE_DIM": 64, "num_stages": 2, "num_warps": 8}
+    )
+    segment_mean_configs.append(
+        {"BLOCK_SIZE_SEQ": 256, "BLOCK_SIZE_HEAD": 8, "BLOCK_SIZE_DIM": 128, "num_stages": 2, "num_warps": 4}
+    )
+
+    if op_name not in OpProtoHolder.instance().op_proto_map.keys():
+        Output = paddle.empty([num_batches, num_heads, head_dim], dtype=x.dtype)
+        prepare_ptr_for_triton_kernel = """
+    // prepare tensor
+    CUdeviceptr input_ptrs[3] = {
+        get_tensor_ptr(x),
+        get_tensor_ptr(output_tensor),
+        get_tensor_ptr(cu_seqlen)
+    };
+"""
+        return_tensor_names = "output_tensor"
+        template_used = rendering_common_template(
+            segment_mean,
+            prepare_attr_for_triton_kernel,
+            prepare_ptr_for_triton_kernel,
+            return_tensor_names,
+        )
+
+        # 确定kernel配置
+        grid = (
+            "num_batches",
+            "(num_heads + BLOCK_SIZE_HEAD - 1) / BLOCK_SIZE_HEAD",
+            "(head_dim + BLOCK_SIZE_DIM - 1) / BLOCK_SIZE_DIM",
+        )
+
+        # 调用kernel
+        segmented_mean_reduce_kernel[(op_name, template_used, grid, segment_mean_configs)](
+            input_ptr=x,
+            output_ptr=Output,
+            cu_seqlen_ptr=cu_seqlen,
+            num_batches=-1,
+            num_heads=num_heads,
+            head_dim=head_dim,
+            input_stride_seq=input_stride_seq,
+            input_stride_head=input_stride_head,
+            output_stride_batch=output_stride_batch,
+            output_stride_head=output_stride_head,
+        )
+
+    if in_dynamic_or_pir_mode():
+        outs = _C_ops._run_custom_op(op_name, x, cu_seqlen)
+        return outs[0]
+    else:
+        helper = LayerHelper(op_name, **locals())
+        inputs = {
+            "x": x,
+            "cu_seqlen_tensor": cu_seqlen,
+        }
+        output = helper.create_variable_for_type_inference(dtype=x.dtype)
+
+        helper.append_op(type=op_name, inputs=inputs, outputs={"output_tensor": output})
+        return output