002f9879b2
Replaces cherry-picked upstream_ref with complete source trees. xllm/ — Iluvatar official C++ inference engine (15MB, 1470 files) Complete: kernels → layers → models → runtime → scheduler → api Excluded: .git, binary images, third_party submodule checkouts ds_vllm/ — Iluvatar official vllm fork (8MB, 703 files) Included: csrc/ (ALL CUDA kernels), fused_moe/, qwen3_5 model, _custom_ops Excluded: tests, benchmarks, docs, examples (not needed for reference) Critical call chains now fully traceable: MoE: moe_topk_softmax_kernels.cuh → ixformer.h → fused_moe.cpp → layer GDN: qwen3_gated_delta_net_base.cpp → qwen3_5_gated_delta_net.cpp Attention: ixformer.h → xllm_paged_attention → attention.cpp
61 lines
1.9 KiB
Plaintext
61 lines
1.9 KiB
Plaintext
#pragma once
|
|
|
|
/**
|
|
* Quantization utilities including:
|
|
* Adjusted maximum values for qtypes.
|
|
* Minimum scaling factors for qtypes.
|
|
*/
|
|
|
|
#include <cmath>
|
|
#include <torch/headeronly/macros/Macros.h>
|
|
|
|
#ifndef USE_ROCM
|
|
#include <torch/headeronly/util/Float8_e4m3fn.h>
|
|
#define MAYBE_HOST_DEVICE C10_HOST_DEVICE
|
|
#else
|
|
#include <torch/headeronly/util/Float8_e4m3fn.h>
|
|
#include <torch/headeronly/util/Float8_e4m3fnuz.h>
|
|
// ROCm doesn't seem to need C10_HOST_DEVICE for static constexpr
|
|
#define MAYBE_HOST_DEVICE
|
|
#endif
|
|
|
|
template <typename T,
|
|
typename = std::enable_if_t<
|
|
std::is_same_v<T, torch::headeronly::Float8_e4m3fn> ||
|
|
std::is_same_v<T, torch::headeronly::Float8_e4m3fnuz> ||
|
|
std::is_same_v<T, int8_t>>>
|
|
struct quant_type_max {
|
|
static constexpr T val() { return std::numeric_limits<T>::max(); }
|
|
};
|
|
|
|
// Using the default max value from pytorch (240.0 0x7F) will cause accuracy
|
|
// issues when running dynamic quantization. Here use 224.0 0x7E for rocm.
|
|
template <>
|
|
struct quant_type_max<torch::headeronly::Float8_e4m3fnuz> {
|
|
static constexpr torch::headeronly::Float8_e4m3fnuz val() {
|
|
return torch::headeronly::Float8_e4m3fnuz(
|
|
0x7E, torch::headeronly::Float8_e4m3fnuz::from_bits());
|
|
}
|
|
};
|
|
|
|
template <typename T>
|
|
MAYBE_HOST_DEVICE static constexpr T quant_type_max_v =
|
|
quant_type_max<T>::val();
|
|
|
|
template <typename T,
|
|
typename = std::enable_if_t<
|
|
std::is_same_v<T, torch::headeronly::Float8_e4m3fn> ||
|
|
std::is_same_v<T, torch::headeronly::Float8_e4m3fnuz> ||
|
|
std::is_same_v<T, int8_t>>>
|
|
struct min_scaling_factor {
|
|
C10_DEVICE C10_ALWAYS_INLINE static float val() {
|
|
return 1.0f / (quant_type_max_v<T> * 512.0f);
|
|
}
|
|
};
|
|
|
|
template <>
|
|
struct min_scaling_factor<int8_t> {
|
|
C10_DEVICE C10_ALWAYS_INLINE static float val() {
|
|
return std::numeric_limits<float>::epsilon();
|
|
}
|
|
}; |