flashinfer-python 0.2.2__tar.gz → 0.2.3__tar.gz

flashinfer_python-0.2.3/PKG-INFO

@@ -0,0 +1,198 @@
+Metadata-Version: 2.2
+Name: flashinfer-python
+Version: 0.2.3
+Summary: FlashInfer: Kernel Library for LLM Serving
+Author: FlashInfer team
+License: Apache License 2.0
+Project-URL: Homepage, https://github.com/flashinfer-ai/flashinfer
+Requires-Python: <4.0,>=3.8
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy
+Requires-Dist: torch
+Requires-Dist: ninja
+Dynamic: requires-dist
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://github.com/flashinfer-ai/web-data/blob/main/logo/FlashInfer-black-background.png?raw=true">
+    <img alt="FlashInfer" src="https://github.com/flashinfer-ai/web-data/blob/main/logo/FlashInfer-white-background.png?raw=true" width=55%>
+  </picture>
+</p>
+<h1 align="center">
+Kernel Library for LLM Serving
+</h1>
+
+<p align="center">
+| <a href="https://flashinfer.ai"><b>Blog</b></a> | <a href="https://docs.flashinfer.ai"><b>Documentation</b></a> | <a href="https://join.slack.com/t/flashinfer/shared_invite/zt-2r93kj2aq-wZnC2n_Z2~mf73N5qnVGGA"><b>Slack</b></a>|  <a href="https://github.com/orgs/flashinfer-ai/discussions"><b>Discussion Forum</b></a> |
+</p>
+
+[![Release](https://github.com/flashinfer-ai/flashinfer/actions/workflows/release_wheel.yml/badge.svg)](https://github.com/flashinfer-ai/flashinfer/actions/workflows/release_wheel.yml)
+[![Documentation](https://github.com/flashinfer-ai/flashinfer/actions/workflows/build-doc.yml/badge.svg)](https://github.com/flashinfer-ai/flashinfer/actions/workflows/build-doc.yml)
+
+
+FlashInfer is a library and kernel generator for Large Language Models that provides high-performance implementation of LLM GPU kernels such as FlashAttention, SparseAttention, PageAttention, Sampling, and more. FlashInfer focuses on LLM serving and inference, and delivers state-of-the-art performance across diverse scenarios.
+
+Check our [v0.2 release blog](https://flashinfer.ai/2024/12/16/flashinfer-v02-release.html) for new features!
+
+The core features of FlashInfer include:
+1. **Efficient Sparse/Dense Attention Kernels**: Efficient single/batch attention for sparse(paged)/dense KV-storage on CUDA Cores and Tensor Cores (both FA2 & FA3) templates. The vector-sparse attention can achieve 90% of the bandwidth of dense kernels with same problem size.
+2. **Load-Balanced Scheduling**: FlashInfer decouples `plan`/`run` stage of attention computation where we schedule the computation of variable-length inputs in `plan` stage to alleviate load-imbalance issue.
+3. **Memory Efficiency**: FlashInfer offers [Cascade Attention](https://docs.flashinfer.ai/api/cascade.html#flashinfer.cascade.MultiLevelCascadeAttentionWrapper) for hierarchical KV-Cache, and implements Head-Query fusion for accelerating Grouped-Query Attention, and efficient kernels for low-precision attention and fused-RoPE attention for compressed KV-Cache.
+4. **Customizable Attention**: Bring your own attention variants through JIT-compilation.
+5. **CUDAGraph and torch.compile Compatibility**: FlashInfer kernels can be captured by CUDAGraphs and torch.compile for low-latency inference.
+6. **Efficient LLM-specific Operators**: High-Performance [fused kernel for Top-P, Top-K/Min-P sampling](https://docs.flashinfer.ai/api/sampling.html) without the need to sorting.
+
+FlashInfer supports PyTorch, TVM and C++ (header-only) APIs, and can be easily integrated into existing projects.
+
+## News
+- [Dec 16, 2024] [Blog Post](https://flashinfer.ai/2024/12/16/flashinfer-v02-release.html) FlashInfer 0.2 - Efficient and Customizable Kernels for LLM Inference Serving
+- [Sept 2024] We've launched a [Slack](https://join.slack.com/t/flashinfer/shared_invite/zt-2r93kj2aq-wZnC2n_Z2~mf73N5qnVGGA) workspace for Flashinfer users and developers. Join us for timely support, discussions, updates and knowledge sharing!
+- [Jan 31, 2024] [Blog Post](https://flashinfer.ai/2024/01/08/cascade-inference.html) Cascade Inference: Memory-Efficient Shared Prefix Batch Decoding
+- [Jan 31, 2024] [Blog Post](https://flashinfer.ai/2024/01/03/introduce-flashinfer.html) Accelerating Self-Attentions for LLM Serving with FlashInfer
+
+## Getting Started
+
+Using our PyTorch API is the easiest way to get started:
+
+### Install from PIP
+
+We provide prebuilt python wheels for Linux. Install FlashInfer with the following command:
+
+```bash
+# For CUDA 12.4 & torch 2.5
+pip install flashinfer-python -i https://flashinfer.ai/whl/cu124/torch2.5
+# For other CUDA & torch versions, check https://docs.flashinfer.ai/installation.html
+```
+
+To try the latest features from the main branch, use our nightly-built wheels:
+
+```bash
+pip install flashinfer-python -i https://flashinfer.ai/whl/nightly/cu124/torch2.4
+```
+
+For a JIT version (compiling every kernel from scratch, [NVCC](https://developer.nvidia.com/cuda-downloads) is required), install from [PyPI](https://pypi.org/project/flashinfer-python/):
+
+```bash
+pip install flashinfer-python
+```
+
+### Install from Source
+
+Alternatively, build FlashInfer from source:
+
+```bash
+git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
+cd flashinfer
+pip install -e . -v
+```
+
+To pre-compile essential kernels, set the environment variable `FLASHINFER_ENABLE_AOT=1` before running the installation command:
+
+```bash
+FLASHINFER_ENABLE_AOT=1 pip install -e . -v
+```
+
+For more details, refer to the [Install from Source documentation](https://docs.flashinfer.ai/installation.html#install-from-source).
+
+### Trying it out
+
+Below is a minimal example of using FlashInfer's single-request decode/append/prefill attention kernels:
+
+```python
+import torch
+import flashinfer
+
+kv_len = 2048
+num_kv_heads = 32
+head_dim = 128
+
+k = torch.randn(kv_len, num_kv_heads, head_dim).half().to(0)
+v = torch.randn(kv_len, num_kv_heads, head_dim).half().to(0)
+
+# decode attention
+
+num_qo_heads = 32
+q = torch.randn(num_qo_heads, head_dim).half().to(0)
+
+o = flashinfer.single_decode_with_kv_cache(q, k, v) # decode attention without RoPE on-the-fly
+o_rope_on_the_fly = flashinfer.single_decode_with_kv_cache(q, k, v, pos_encoding_mode="ROPE_LLAMA") # decode with LLaMA style RoPE on-the-fly
+
+# append attention
+append_qo_len = 128
+q = torch.randn(append_qo_len, num_qo_heads, head_dim).half().to(0) # append attention, the last 128 tokens in the KV-Cache are the new tokens
+o = flashinfer.single_prefill_with_kv_cache(q, k, v, causal=True) # append attention without RoPE on-the-fly, apply causal mask
+o_rope_on_the_fly = flashinfer.single_prefill_with_kv_cache(q, k, v, causal=True, pos_encoding_mode="ROPE_LLAMA") # append attention with LLaMA style RoPE on-the-fly, apply causal mask
+
+# prefill attention
+qo_len = 2048
+q = torch.randn(qo_len, num_qo_heads, head_dim).half().to(0) # prefill attention
+o = flashinfer.single_prefill_with_kv_cache(q, k, v, causal=False) # prefill attention without RoPE on-the-fly, do not apply causal mask
+```
+
+Check out [documentation](https://docs.flashinfer.ai/) for usage of batch decode/append/prefill kernels and shared-prefix cascading kernels.
+
+## Custom Attention Variants
+
+Starting from FlashInfer v0.2, users can customize their own attention variants with additional parameters. For more details, refer to our [JIT examples](https://github.com/flashinfer-ai/flashinfer/blob/main/tests/test_jit_example.py).
+
+## Run Benchmarks
+
+We profile FlashInfer kernel performance with [nvbench](https://github.com/NVIDIA/nvbench) and you can compile and run the benchmarks with the following commands:
+
+```bash
+mkdir build
+cp cmake/config.cmake build # you can modify the config.cmake to enable/disable benchmarks and change CUDA architectures
+cd build
+cmake ..
+make -j12
+```
+
+You can run `./bench_{single/batch}_{prefill/decode}` to benchmark the performance (e.g. `./bench_single_prefill` for single-request prefill attention). `./bench_{single/batch}_{prefill/decode} --help` will show you the available options.
+
+## C++ API and TVM Bindings
+
+FlashInfer also provides C++ API and TVM bindings, please refer to [documentation](https://docs.flashinfer.ai/) for more details.
+
+## Adoption
+
+We are thrilled to share that FlashInfer is being adopted by many cutting-edge projects, including but not limited to:
+- [MLC-LLM](https://github.com/mlc-ai/mlc-llm)
+- [Punica](https://github.com/punica-ai/punica)
+- [SGLang](https://github.com/sgl-project/sglang)
+- [ScaleLLM](https://github.com/vectorch-ai/ScaleLLM)
+- [vLLM](https://github.com/vllm-project/vllm)
+- [TGI](https://github.com/huggingface/text-generation-inference)
+- [lorax](https://github.com/predibase/lorax)
+- [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM)
+- [LightLLM](https://github.com/ModelTC/lightllm)
+
+## Acknowledgement
+
+FlashInfer is inspired by [FlashAttention 1&2](https://github.com/dao-AILab/flash-attention/), [vLLM](https://github.com/vllm-project/vllm), [stream-K](https://arxiv.org/abs/2301.03598), [cutlass](https://github.com/nvidia/cutlass) and [AITemplate](https://github.com/facebookincubator/AITemplate) projects.
+
+## Citation
+
+If you find FlashInfer helpful in your project or research, please consider citing our [paper](https://arxiv.org/abs/2501.01005):
+
+```bibtex
+@article{ye2025flashinfer,
+    title = {FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving},
+    author = {
+      Ye, Zihao and
+      Chen, Lequn and
+      Lai, Ruihang and
+      Lin, Wuwei and
+      Zhang, Yineng and
+      Wang, Stephanie and
+      Chen, Tianqi and
+      Kasikci, Baris and
+      Grover, Vinod and
+      Krishnamurthy, Arvind and
+      Ceze, Luis
+    },
+    journal = {arXiv preprint arXiv:2501.01005},
+    year = {2025},
+    url = {https://arxiv.org/abs/2501.01005}
+}
+```

flashinfer_python-0.2.3/csrc/activation.cu

@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2024 by FlashInfer team.
+ *
+ * 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.
+ */
+#include <flashinfer/activation.cuh>
+
+#include "pytorch_extension_utils.h"
+
+using namespace flashinfer;
+
+__device__ __forceinline__ float silu(const float& val) { return val / (1.0f + __expf(-val)); }
+
+__device__ __forceinline__ float gelu(const float& val) {
+  constexpr float kAlpha = M_SQRT1_2;
+  return val * 0.5f * (1.0f + ::erf(val * kAlpha));
+}
+
+__device__ __forceinline__ float gelu_tanh(const float& val) {
+  const float cdf =
+      0.5f * (1.0f + math::tanh((0.7978845608028654f * (val + 0.044715f * val * val * val))));
+  return val * cdf;
+}
+
+void silu_and_mul(at::Tensor& out, at::Tensor& input, bool enable_pdl, int64_t cuda_stream) {
+  int d = input.size(-1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input.scalar_type(), c_type, [&] {
+    uint32_t vec_size = 16 / sizeof(c_type);
+    cudaLaunchConfig_t config;
+    config.gridDim = num_tokens;
+    config.blockDim = std::min(d / vec_size, 1024U);
+    config.dynamicSmemBytes = 0;
+    config.stream = stream;
+    cudaLaunchAttribute attrs[1];
+    attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+    attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
+    config.numAttrs = 1;
+    config.attrs = attrs;
+
+    auto kernel = flashinfer::activation::act_and_mul_kernel<c_type, silu>;
+
+    cudaLaunchKernelEx(&config, kernel, static_cast<c_type*>(out.data_ptr()),
+                       static_cast<c_type*>(input.data_ptr()), d);
+
+    cudaError_t err = cudaGetLastError();
+    TORCH_CHECK(err == cudaSuccess, "Failed to launch kernel: ", cudaGetErrorString(err));
+
+    return true;
+  });
+}
+
+void gelu_tanh_and_mul(at::Tensor& out, at::Tensor& input, bool enable_pdl, int64_t cuda_stream) {
+  int d = input.size(-1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input.scalar_type(), c_type, [&] {
+    uint32_t vec_size = 16 / sizeof(c_type);
+    cudaLaunchConfig_t config;
+    config.gridDim = num_tokens;
+    config.blockDim = std::min(d / vec_size, 1024U);
+    config.dynamicSmemBytes = 0;
+    config.stream = stream;
+    cudaLaunchAttribute attrs[1];
+    attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+    attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
+    config.numAttrs = 1;
+    config.attrs = attrs;
+
+    auto kernel = flashinfer::activation::act_and_mul_kernel<c_type, gelu_tanh>;
+
+    cudaLaunchKernelEx(&config, kernel, static_cast<c_type*>(out.data_ptr()),
+                       static_cast<c_type*>(input.data_ptr()), d);
+
+    cudaError_t err = cudaGetLastError();
+    TORCH_CHECK(err == cudaSuccess, "Failed to launch kernel: ", cudaGetErrorString(err));
+
+    return true;
+  });
+}
+
+void gelu_and_mul(at::Tensor& out, at::Tensor& input, bool enable_pdl, int64_t cuda_stream) {
+  int d = input.size(-1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+  dim3 grid(num_tokens);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input.scalar_type(), c_type, [&] {
+    uint32_t vec_size = 16 / sizeof(c_type);
+    cudaLaunchConfig_t config;
+    config.gridDim = num_tokens;
+    config.blockDim = std::min(d / vec_size, 1024U);
+    config.dynamicSmemBytes = 0;
+    config.stream = stream;
+    cudaLaunchAttribute attrs[1];
+    attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+    attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
+    config.numAttrs = 1;
+    config.attrs = attrs;
+
+    auto kernel = flashinfer::activation::act_and_mul_kernel<c_type, gelu>;
+
+    cudaLaunchKernelEx(&config, kernel, static_cast<c_type*>(out.data_ptr()),
+                       static_cast<c_type*>(input.data_ptr()), d);
+
+    cudaError_t err = cudaGetLastError();
+    TORCH_CHECK(err == cudaSuccess, "Failed to launch kernel: ", cudaGetErrorString(err));
+
+    return true;
+  });
+}

flashinfer_python-0.2.3/csrc/batch_mla_config.jinja

@@ -0,0 +1,33 @@
+#pragma once
+#include <flashinfer/page.cuh>
+#include <flashinfer/math.cuh>
+#include <flashinfer/layout.cuh>
+#include <flashinfer/utils.cuh>
+#include <flashinfer/pos_enc.cuh>
+#include <flashinfer/fastdiv.cuh>
+#include <flashinfer/attention/variant_helper.cuh>
+#include <flashinfer/attention/mla_params.cuh>
+
+using namespace flashinfer;
+
+#ifdef FLASHINFER_ENABLE_PROFILER
+#define ADDITIONAL_FUNC_PARAMS , at::Tensor profiler_buffer
+#define ADDITIONAL_PARAMS_SETTER \
+  params.profiler_buffer = static_cast<uint64_t*>(profiler_buffer.data_ptr());
+#else
+#define ADDITIONAL_FUNC_PARAMS
+#define ADDITIONAL_PARAMS_SETTER
+#endif
+
+using DTypeQ = {{ dtype_q }};
+using DTypeKV = {{ dtype_kv }};
+using DTypeO = {{ dtype_o }};
+using IdType = {{ dtype_idx }};
+constexpr int HEAD_DIM_CKV = {{ head_dim_ckv }};
+constexpr int HEAD_DIM_KPE = {{ head_dim_kpe }};
+
+#define DISPATCH_context(DTypeQ, DTypeKV, DTypeO, IdType, MASK_MODE, HEAD_DIM_CKV, HEAD_DIM_KPE, Params, ...) \
+  DISPATCH_MASK_MODE(mask_mode, MASK_MODE, { \
+    using Params = MLAParams<DTypeQ, DTypeKV, DTypeO, IdType>; \
+    __VA_ARGS__(); \
+  })

flashinfer_python-0.2.3/csrc/batch_mla_run.cu

@@ -0,0 +1,124 @@
+/*
+ * Copyright (c) 2025 by FlashInfer team.
+ *
+ * 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.
+ */
+#include <flashinfer/attention/mla.cuh>
+#include <flashinfer/attention/scheduler.cuh>
+#include <flashinfer/fastdiv.cuh>
+#include <optional>
+
+#include "batch_mla_config.inc"
+#include "pytorch_conversion_utils.h"
+#include "pytorch_extension_utils.h"
+
+using namespace flashinfer;
+
+void BatchMLAPagedAttentionRun(at::Tensor float_workspace_buffer, at::Tensor int_workspace_buffer,
+                               at::Tensor plan_info_vec, at::Tensor q_nope, at::Tensor q_pe,
+                               at::Tensor ckv_cache, at::Tensor kpe_cache, at::Tensor kv_indices,
+                               at::Tensor o, std::optional<at::Tensor> maybe_lse,
+                               int64_t mask_mode_code, int64_t num_heads, int64_t page_size,
+                               double sm_scale, int64_t cuda_stream) {
+  // q_nope: [n, num_heads, head_dim_ckv]
+  // q_pe: [n, num_heads, head_dim_kpe]
+  // ckv_cache: [num_pages, page_size, head_dim_ckv]
+  // kpe_cache: [num_pages, page_size, head_dim_kpe]
+  MLAPlanInfo plan_info;
+  plan_info.FromVector(tensor_to_vec(plan_info_vec));
+
+  auto device = q_nope.device();
+
+  void* float_buffer_ptr = float_workspace_buffer.data_ptr();
+  void* int_buffer_ptr = int_workspace_buffer.data_ptr();
+
+  const MaskMode mask_mode = static_cast<MaskMode>(mask_mode_code);
+
+  auto q_scalar_type = q_nope.scalar_type();
+  auto kv_scalar_type = ckv_cache.scalar_type();
+
+  unsigned int q_nope_stride_n = q_nope.stride(0);
+  unsigned int q_nope_stride_h = q_nope.stride(1);
+  unsigned int q_pe_stride_n = q_pe.stride(0);
+  unsigned int q_pe_stride_h = q_pe.stride(1);
+  unsigned int ckv_stride_page = ckv_cache.stride(0);
+  unsigned int ckv_stride_n = ckv_cache.stride(1);
+  unsigned int kpe_stride_page = kpe_cache.stride(0);
+  unsigned int kpe_stride_n = kpe_cache.stride(1);
+  unsigned int o_stride_n = o.stride(0);
+  unsigned int o_stride_h = o.stride(1);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+
+  DISPATCH_context(
+      DTypeQ, DTypeKV, DTypeO, IdType, MASK_MODE, HEAD_DIM_CKV, HEAD_DIM_KPE, Params, [&] {
+        Params params;
+
+        params.q_nope = static_cast<DTypeQ*>(q_nope.data_ptr());
+        params.q_pe = static_cast<DTypeQ*>(q_pe.data_ptr());
+        params.ckv = static_cast<DTypeKV*>(ckv_cache.data_ptr());
+        params.kpe = static_cast<DTypeKV*>(kpe_cache.data_ptr());
+
+        params.q_indptr = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.q_indptr_offset);
+        params.kv_indptr = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_indptr_offset);
+        params.partial_indptr =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.partial_indptr_offset);
+        params.kv_indices = static_cast<IdType*>(kv_indices.data_ptr());
+        params.q_len = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.q_len_offset);
+        params.kv_len = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_len_offset);
+        params.q_start = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.q_start_offset);
+        params.kv_start = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_start_offset);
+        params.kv_end = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_end_offset);
+        params.work_indptr =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.work_indptr_offset);
+        params.merge_packed_offset_start = GetPtrFromBaseOffset<IdType>(
+            int_buffer_ptr, plan_info.merge_packed_offset_start_offset);
+        params.merge_packed_offset_end =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.merge_packed_offset_end_offset);
+        params.merge_partial_packed_offset_start = GetPtrFromBaseOffset<IdType>(
+            int_buffer_ptr, plan_info.merge_partial_packed_offset_start_offset);
+        params.merge_partial_packed_offset_end = GetPtrFromBaseOffset<IdType>(
+            int_buffer_ptr, plan_info.merge_partial_packed_offset_end_offset);
+        params.merge_partial_stride =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.merge_partial_stride_offset);
+        params.final_o = static_cast<DTypeO*>(o.data_ptr());
+        params.final_lse =
+            maybe_lse.has_value() ? static_cast<float*>(maybe_lse->data_ptr()) : nullptr;
+        params.partial_o =
+            GetPtrFromBaseOffset<DTypeO>(float_buffer_ptr, plan_info.partial_o_offset);
+        params.partial_lse =
+            GetPtrFromBaseOffset<float>(float_buffer_ptr, plan_info.partial_lse_offset);
+
+        params.num_heads = uint_fastdiv(num_heads);
+        params.block_size = uint_fastdiv(page_size);
+
+        params.q_nope_stride_n = q_nope_stride_n;
+        params.q_nope_stride_h = q_nope_stride_h;
+        params.q_pe_stride_n = q_pe_stride_n;
+        params.q_pe_stride_h = q_pe_stride_h;
+        params.ckv_stride_page = ckv_stride_page;
+        params.ckv_stride_n = ckv_stride_n;
+        params.kpe_stride_page = kpe_stride_page;
+        params.kpe_stride_n = kpe_stride_n;
+        params.o_stride_n = o_stride_n;
+        params.o_stride_h = o_stride_h;
+
+        params.sm_scale = sm_scale;
+
+        cudaError_t status = mla::BatchMLAPagedAttention<MASK_MODE, HEAD_DIM_CKV, HEAD_DIM_KPE>(
+            params, plan_info.num_blks_x, plan_info.num_blks_y, stream);
+
+        TORCH_CHECK(status == cudaSuccess,
+                    "Failed to run MLA, error: ", cudaGetErrorString(status));
+      });
+}

flashinfer_python-0.2.3/csrc/batch_mla_sm90_pybind.cu

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2025 by FlashInfer team.
+ *
+ * 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.
+ */
+#include "batch_mla_sm90_config.inc"
+#include "pytorch_extension_utils.h"
+
+at::Tensor BatchMLAPagedAttentionSM90Plan(at::Tensor float_workspace_buffer,
+                                          at::Tensor int_workspace_buffer,
+                                          at::Tensor page_locked_int_workspace_buffer,
+                                          at::Tensor qo_indptr, at::Tensor kv_indptr,
+                                          at::Tensor kv_len, int64_t num_heads, int64_t head_dim_o,
+                                          bool causal, int64_t cuda_stream);
+
+void BatchMLAPagedAttentionSM90Run(at::Tensor float_workspace_buffer,
+                                   at::Tensor int_workspace_buffer, at::Tensor plan_info_vec,
+                                   at::Tensor q_nope, at::Tensor q_pe, at::Tensor ckv_cache,
+                                   at::Tensor kpe_cache, at::Tensor kv_indices, at::Tensor o,
+                                   std::optional<at::Tensor> maybe_lse, int64_t mask_mode_code,
+                                   int64_t num_heads, int64_t page_size,
+                                   double sm_scale ADDITIONAL_FUNC_PARAMS, int64_t cuda_stream);
+
+TORCH_LIBRARY_FRAGMENT(TORCH_EXTENSION_NAME, m) {
+  m.def("plan", &BatchMLAPagedAttentionSM90Plan);
+  m.def("run", &BatchMLAPagedAttentionSM90Run);
+}

flashinfer_python-0.2.3/csrc/batch_mla_sm90_run.cu

@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2025 by FlashInfer team.
+ *
+ * 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.
+ */
+#include <flashinfer/attention/mla_hopper.cuh>
+#include <flashinfer/attention/scheduler.cuh>
+#include <flashinfer/fastdiv.cuh>
+#include <optional>
+
+#include "batch_mla_sm90_config.inc"
+#include "pytorch_conversion_utils.h"
+#include "pytorch_extension_utils.h"
+
+using namespace flashinfer;
+
+void BatchMLAPagedAttentionSM90Run(at::Tensor float_workspace_buffer,
+                                   at::Tensor int_workspace_buffer, at::Tensor plan_info_vec,
+                                   at::Tensor q_nope, at::Tensor q_pe, at::Tensor ckv_cache,
+                                   at::Tensor kpe_cache, at::Tensor kv_indices, at::Tensor o,
+                                   std::optional<at::Tensor> maybe_lse, int64_t mask_mode_code,
+                                   int64_t num_heads, int64_t page_size,
+                                   double sm_scale ADDITIONAL_FUNC_PARAMS, int64_t cuda_stream) {
+  // q_nope: [n, num_heads, head_dim_ckv]
+  // q_pe: [n, num_heads, head_dim_kpe]
+  // ckv_cache: [num_pages, page_size, head_dim_ckv]
+  // kpe_cache: [num_pages, page_size, head_dim_kpe]
+  MLAPlanInfo plan_info;
+  plan_info.FromVector(tensor_to_vec(plan_info_vec));
+
+  auto device = q_nope.device();
+
+  void* float_buffer_ptr = float_workspace_buffer.data_ptr();
+  void* int_buffer_ptr = int_workspace_buffer.data_ptr();
+
+  const MaskMode mask_mode = static_cast<MaskMode>(mask_mode_code);
+
+  auto q_scalar_type = q_nope.scalar_type();
+  auto kv_scalar_type = ckv_cache.scalar_type();
+
+  unsigned int q_nope_stride_n = q_nope.stride(0);
+  unsigned int q_nope_stride_h = q_nope.stride(1);
+  unsigned int q_pe_stride_n = q_pe.stride(0);
+  unsigned int q_pe_stride_h = q_pe.stride(1);
+  unsigned int ckv_stride_page = ckv_cache.stride(0);
+  unsigned int ckv_stride_n = ckv_cache.stride(1);
+  unsigned int kpe_stride_page = kpe_cache.stride(0);
+  unsigned int kpe_stride_n = kpe_cache.stride(1);
+  unsigned int o_stride_n = o.stride(0);
+  unsigned int o_stride_h = o.stride(1);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+
+  DISPATCH_context(
+      DTypeQ, DTypeKV, DTypeO, IdType, MASK_MODE, HEAD_DIM_CKV, HEAD_DIM_KPE, Params, [&] {
+        Params params;
+
+        params.q_nope = static_cast<DTypeQ*>(q_nope.data_ptr());
+        params.q_pe = static_cast<DTypeQ*>(q_pe.data_ptr());
+        params.ckv = static_cast<DTypeKV*>(ckv_cache.data_ptr());
+        params.kpe = static_cast<DTypeKV*>(kpe_cache.data_ptr());
+
+        params.q_indptr = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.q_indptr_offset);
+        params.kv_indptr = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_indptr_offset);
+        params.partial_indptr =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.partial_indptr_offset);
+        params.kv_indices = static_cast<IdType*>(kv_indices.data_ptr());
+        params.q_len = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.q_len_offset);
+        params.kv_len = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_len_offset);
+        params.q_start = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.q_start_offset);
+        params.kv_start = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_start_offset);
+        params.kv_end = GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.kv_end_offset);
+        params.work_indptr =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.work_indptr_offset);
+        params.merge_packed_offset_start = GetPtrFromBaseOffset<IdType>(
+            int_buffer_ptr, plan_info.merge_packed_offset_start_offset);
+        params.merge_packed_offset_end =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.merge_packed_offset_end_offset);
+        params.merge_partial_packed_offset_start = GetPtrFromBaseOffset<IdType>(
+            int_buffer_ptr, plan_info.merge_partial_packed_offset_start_offset);
+        params.merge_partial_packed_offset_end = GetPtrFromBaseOffset<IdType>(
+            int_buffer_ptr, plan_info.merge_partial_packed_offset_end_offset);
+        params.merge_partial_stride =
+            GetPtrFromBaseOffset<IdType>(int_buffer_ptr, plan_info.merge_partial_stride_offset);
+        params.final_o = static_cast<DTypeO*>(o.data_ptr());
+        params.final_lse =
+            maybe_lse.has_value() ? static_cast<float*>(maybe_lse->data_ptr()) : nullptr;
+        params.partial_o =
+            GetPtrFromBaseOffset<DTypeO>(float_buffer_ptr, plan_info.partial_o_offset);
+        params.partial_lse =
+            GetPtrFromBaseOffset<float>(float_buffer_ptr, plan_info.partial_lse_offset);
+
+        params.num_heads = uint_fastdiv(num_heads);
+        params.block_size = uint_fastdiv(page_size);
+
+        params.q_nope_stride_n = q_nope_stride_n;
+        params.q_nope_stride_h = q_nope_stride_h;
+        params.q_pe_stride_n = q_pe_stride_n;
+        params.q_pe_stride_h = q_pe_stride_h;
+        params.ckv_stride_page = ckv_stride_page;
+        params.ckv_stride_n = ckv_stride_n;
+        params.kpe_stride_page = kpe_stride_page;
+        params.kpe_stride_n = kpe_stride_n;
+        params.o_stride_n = o_stride_n;
+        params.o_stride_h = o_stride_h;
+
+        ADDITIONAL_PARAMS_SETTER
+
+        params.sm_scale = sm_scale;
+
+        cudaError_t status =
+            mla::BatchMLAPageAttentionHopper<MASK_MODE, HEAD_DIM_CKV, HEAD_DIM_KPE>(
+                params, plan_info.num_blks_x, plan_info.num_blks_y, stream);
+
+        TORCH_CHECK(status == cudaSuccess,
+                    "Failed to run MLA, error: ", cudaGetErrorString(status));
+      });
+}