forked from apache/tvm
-
Notifications
You must be signed in to change notification settings - Fork 2
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge pull request apache#61 from octoml/mlc-serve-v0.2.0-feature-fp8
[FP8] Bring fp8 support to OLLM tracking branch
- Loading branch information
Showing
6 changed files
with
219 additions
and
1 deletion.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,135 @@ | ||
| /* | ||
| * Licensed to the Apache Software Foundation (ASF) under one | ||
| * or more contributor license agreements. See the NOTICE file | ||
| * distributed with this work for additional information | ||
| * regarding copyright ownership. The ASF licenses this file | ||
| * to you 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. | ||
| */ | ||
|
|
||
| /*! | ||
| * \file Externally defined CUDA kernels for use in TVM runtime | ||
| */ | ||
|
|
||
| #include <cuda_fp16.h> | ||
| #include <cuda_runtime.h> | ||
| #include <dlpack/dlpack.h> | ||
| #include <tvm/runtime/registry.h> | ||
|
|
||
| #include "../../cuda/cuda_common.h" | ||
|
|
||
| namespace tvm { | ||
| namespace contrib { | ||
|
|
||
| using namespace runtime; | ||
|
|
||
| template <typename T> | ||
| __device__ T device_max(T a, T b) { | ||
| return max(a, b); | ||
| } | ||
|
|
||
| template <> | ||
| __device__ __half device_max(__half a, __half b) { | ||
| return __hmax(a, b); | ||
| } | ||
|
|
||
| template <typename T> | ||
| __device__ T device_abs(T a) { | ||
| return abs(a); | ||
| } | ||
|
|
||
| template <> | ||
| __device__ __half device_abs(__half a) { | ||
| return __habs(a); | ||
| } | ||
|
|
||
| template <typename T> | ||
| __inline__ __device__ T warp_reduce_max(T val) { | ||
| for (int offset = warpSize / 2; offset > 0; offset >>= 1) { | ||
| val = device_max(val, __shfl_down_sync(0xffffffff, val, offset)); | ||
| } | ||
| return val; | ||
| } | ||
|
|
||
| // Single block reduce, assumes size % 1024 == 0 | ||
| template <typename T> | ||
| __global__ void max_reduce_kernel_single_block(T* input, T* output, int size) { | ||
| __shared__ T shared[32]; | ||
|
|
||
| int tid = threadIdx.x; | ||
| T max_val = std::numeric_limits<T>::lowest(); | ||
|
|
||
| // Step 1: Each thread reduces across the elements it owns | ||
| for (int i = tid; i < size; i += blockDim.x) { | ||
| // use __hmax for float16 | ||
| max_val = device_max(max_val, device_abs(input[i])); | ||
| } | ||
|
|
||
| // Step 2: Perform reduce across warps | ||
| max_val = warp_reduce_max(max_val); | ||
|
|
||
| // Step 3: Write the reduced value from each warp to shared memory | ||
| if (tid % warpSize == 0) { | ||
| shared[tid / warpSize] = max_val; | ||
| } | ||
| __syncthreads(); | ||
|
|
||
| // Step 4: Perform a final reduction in the first warp across shared values | ||
| if (tid < warpSize) { | ||
| max_val = shared[tid]; | ||
| max_val = warp_reduce_max(max_val); | ||
| if (tid == 0) { | ||
| *output = max_val; | ||
| } | ||
| } | ||
| } | ||
|
|
||
| template __global__ void max_reduce_kernel_single_block<float>(float* input, float* output, | ||
| int size); | ||
| template __global__ void max_reduce_kernel_single_block<__half>(__half* input, __half* output, | ||
| int size); | ||
| template <typename T> | ||
| void LaunchMaxReduceKernelSingleBlock(DLTensor* input, DLTensor* output, int size) { | ||
| T* input_ptr = static_cast<T*>(input->data); | ||
| T* output_ptr = static_cast<T*>(output->data); | ||
|
|
||
| int blocks = 1; | ||
| int threads = 1024; | ||
| max_reduce_kernel_single_block<T><<<blocks, threads>>>(input_ptr, output_ptr, size); | ||
| } | ||
|
|
||
| TVM_REGISTER_GLOBAL("tvm.contrib.cuda.reduce_max_abs").set_body([](TVMArgs args, TVMRetValue* ret) { | ||
| DLTensor* input = args[0]; | ||
| DLTensor* output = args[1]; | ||
|
|
||
| int size = 1; | ||
| for (int i = 0; i < input->ndim; ++i) { | ||
| size *= input->shape[i]; | ||
| } | ||
|
|
||
| CHECK_EQ(size % 1024, 0) << "tvm.contrib.cuda.reduce_max_abs currently only supports reducing " | ||
| "tensors that are an even factor of 1024 elements"; | ||
|
|
||
| auto dtype = DLDataType2String(input->dtype); | ||
|
|
||
| if (dtype == "float32") { | ||
| LaunchMaxReduceKernelSingleBlock<float>(input, output, size); | ||
| } else if (dtype == "float16") { | ||
| LaunchMaxReduceKernelSingleBlock<__half>(input, output, size); | ||
| } else { | ||
| LOG(FATAL) << "Unsupported input dtype: " << dtype; | ||
| } | ||
| }); | ||
|
|
||
| } // namespace contrib | ||
| } // namespace tvm |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,20 @@ | ||
| import numpy as np | ||
| import tvm | ||
| import tvm.testing | ||
|
|
||
|
|
||
| def test_reduce_max_abs(): | ||
| target = "cuda" | ||
| dev = tvm.device(target, 0) | ||
| x_shape = (4, 4096) | ||
| dtype = "float16" | ||
| x = tvm.nd.array(np.random.uniform(-2, 1.4, x_shape).astype(dtype), dev) | ||
| scalar = tvm.nd.array(np.array([0], dtype=dtype), dev) | ||
|
|
||
| reduce = tvm._ffi.get_global_func("tvm.contrib.cuda.reduce_max_abs") | ||
| reduce(x, scalar) | ||
| tvm.testing.assert_allclose(scalar.numpy(), np.array([2], dtype=dtype)) | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| tvm.testing.main() |