mlx/mlx/fast.h

93 lines
2.1 KiB
C
Raw Normal View History

// Copyright © 2023-2024 Apple Inc.
#pragma once
Fast Inference SDPA op (#735) * Fast Inference SDPA op Implements metal shaders for: o = mx.fast_inference_sdpa(queries, keys, values, scale, mask) Supports fp16, fp32 dtypes; assumes d_k = 128. Generic op support / prompt encoding supported via mlx primitives. Metal implementation is for the inference use case only. Majority of performance benefits appears to results from GQA & reduced bandwidth requirements; there is approximate performance parity for the MHA use case (from some measurements on M3 Max). * Flush shared memory to zero before unprotected reads for (scores @ values) * Move to fast:: namespace, address reviewer comments ... also attempt to revert formatter auto-change for files not relevant to this change * Shared memory flush to top of kernel * Resolve compiler warnings * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update docstring per PR feedback * Softmax in higher precision, ... * route to fallback for more use cases - batch size > 1, head_dim other than 128, etc. * Address linux build failure * Address other reviewer comments * Remove extraneous eval_cpu function per review --------- Co-authored-by: Atila Orhon <64497909+atiorh@users.noreply.github.com> Co-authored-by: Awni Hannun <awni.hannun@gmail.com> Co-authored-by: atila <atiorh@icloud.com>
2024-03-05 13:06:11 +08:00
#include <optional>
#include <variant>
Fast Inference SDPA op (#735) * Fast Inference SDPA op Implements metal shaders for: o = mx.fast_inference_sdpa(queries, keys, values, scale, mask) Supports fp16, fp32 dtypes; assumes d_k = 128. Generic op support / prompt encoding supported via mlx primitives. Metal implementation is for the inference use case only. Majority of performance benefits appears to results from GQA & reduced bandwidth requirements; there is approximate performance parity for the MHA use case (from some measurements on M3 Max). * Flush shared memory to zero before unprotected reads for (scores @ values) * Move to fast:: namespace, address reviewer comments ... also attempt to revert formatter auto-change for files not relevant to this change * Shared memory flush to top of kernel * Resolve compiler warnings * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update docstring per PR feedback * Softmax in higher precision, ... * route to fallback for more use cases - batch size > 1, head_dim other than 128, etc. * Address linux build failure * Address other reviewer comments * Remove extraneous eval_cpu function per review --------- Co-authored-by: Atila Orhon <64497909+atiorh@users.noreply.github.com> Co-authored-by: Awni Hannun <awni.hannun@gmail.com> Co-authored-by: atila <atiorh@icloud.com>
2024-03-05 13:06:11 +08:00
#include "mlx/utils.h"
namespace mlx::core::fast {
array rms_norm(
const array& x,
2025-03-01 12:26:57 +08:00
const std::optional<array>& weight,
float eps,
StreamOrDevice s = {});
2024-03-22 04:55:51 +08:00
array layer_norm(
const array& x,
const std::optional<array>& weight,
const std::optional<array>& bias,
float eps,
StreamOrDevice s = {});
array rope(
const array& x,
int dims,
bool traditional,
std::optional<float> base,
float scale,
int offset,
const std::optional<array>& freqs = std::nullopt,
StreamOrDevice s = {});
array rope(
const array& x,
int dims,
bool traditional,
std::optional<float> base,
float scale,
const array& offset,
const std::optional<array>& freqs = std::nullopt,
StreamOrDevice s = {});
Fast Inference SDPA op (#735) * Fast Inference SDPA op Implements metal shaders for: o = mx.fast_inference_sdpa(queries, keys, values, scale, mask) Supports fp16, fp32 dtypes; assumes d_k = 128. Generic op support / prompt encoding supported via mlx primitives. Metal implementation is for the inference use case only. Majority of performance benefits appears to results from GQA & reduced bandwidth requirements; there is approximate performance parity for the MHA use case (from some measurements on M3 Max). * Flush shared memory to zero before unprotected reads for (scores @ values) * Move to fast:: namespace, address reviewer comments ... also attempt to revert formatter auto-change for files not relevant to this change * Shared memory flush to top of kernel * Resolve compiler warnings * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update docstring per PR feedback * Softmax in higher precision, ... * route to fallback for more use cases - batch size > 1, head_dim other than 128, etc. * Address linux build failure * Address other reviewer comments * Remove extraneous eval_cpu function per review --------- Co-authored-by: Atila Orhon <64497909+atiorh@users.noreply.github.com> Co-authored-by: Awni Hannun <awni.hannun@gmail.com> Co-authored-by: atila <atiorh@icloud.com>
2024-03-05 13:06:11 +08:00
/** Computes: O = softmax(Q @ K.T) @ V **/
array scaled_dot_product_attention(
const array& queries,
const array& keys,
const array& values,
const float scale,
const std::string& mask_mode = "",
const std::vector<array>& mask_arrs = {},
Fast Inference SDPA op (#735) * Fast Inference SDPA op Implements metal shaders for: o = mx.fast_inference_sdpa(queries, keys, values, scale, mask) Supports fp16, fp32 dtypes; assumes d_k = 128. Generic op support / prompt encoding supported via mlx primitives. Metal implementation is for the inference use case only. Majority of performance benefits appears to results from GQA & reduced bandwidth requirements; there is approximate performance parity for the MHA use case (from some measurements on M3 Max). * Flush shared memory to zero before unprotected reads for (scores @ values) * Move to fast:: namespace, address reviewer comments ... also attempt to revert formatter auto-change for files not relevant to this change * Shared memory flush to top of kernel * Resolve compiler warnings * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update python/src/fast.cpp Co-authored-by: Awni Hannun <awni.hannun@gmail.com> * Update docstring per PR feedback * Softmax in higher precision, ... * route to fallback for more use cases - batch size > 1, head_dim other than 128, etc. * Address linux build failure * Address other reviewer comments * Remove extraneous eval_cpu function per review --------- Co-authored-by: Atila Orhon <64497909+atiorh@users.noreply.github.com> Co-authored-by: Awni Hannun <awni.hannun@gmail.com> Co-authored-by: atila <atiorh@icloud.com>
2024-03-05 13:06:11 +08:00
StreamOrDevice s = {});
std::tuple<array, array, array> affine_quantize(
const array& w,
int group_size = 64,
int bits = 4,
StreamOrDevice s = {});
array affine_dequantize(
const array& w,
const array& scales,
const array& biases,
int group_size = 64,
int bits = 4,
StreamOrDevice s = {});
typedef std::variant<int, bool, Dtype> TemplateArg;
typedef std::function<std::vector<array>(
const std::vector<array>&,
const std::vector<Shape>&,
const std::vector<Dtype>&,
std::tuple<int, int, int>,
std::tuple<int, int, int>,
std::vector<std::pair<std::string, TemplateArg>>,
std::optional<float>,
bool,
StreamOrDevice)>
MetalKernelFunction;
MetalKernelFunction metal_kernel(
const std::string& name,
const std::vector<std::string>& input_names,
const std::vector<std::string>& output_names,
const std::string& source,
const std::string& header = "",
bool ensure_row_contiguous = true,
bool atomic_outputs = false);
} // namespace mlx::core::fast