mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-09-23 05:26:19 +00:00
5dc9dd7152
* Add Command R Plus GGUF * Add Command R Plus GGUF * Loading works up to LayerNorm2D * Export new tensors in 1D so they are not quantized. * Fix embedding layer based on Noeda's example * Whitespace * Add line * Fix unexpected tokens on MPS. Re-add F16 fix. ((Noeda) * dranger003: Fix block index overflow in CUDA dequantizing. * Reverted blocked multiplication code as it still has issues and could affect other Llama arches * export norms as f32 * fix overflow issues during quant and other cleanup * Type convention Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * dranger003: Fix more int overflow during quant. --------- Co-authored-by: S <seast@Ss-Mac-Studio.local> Co-authored-by: S <s@example.com> Co-authored-by: slaren <slarengh@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
104 lines
2.6 KiB
Plaintext
104 lines
2.6 KiB
Plaintext
#include "common.cuh"
|
|
|
|
static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
|
|
const block_q4_0 * x = (const block_q4_0 *) vx;
|
|
|
|
const dfloat d = x[ib].d;
|
|
|
|
const int vui = x[ib].qs[iqs];
|
|
|
|
v.x = vui & 0xF;
|
|
v.y = vui >> 4;
|
|
|
|
#ifdef GGML_CUDA_F16
|
|
v = __hsub2(v, {8.0f, 8.0f});
|
|
v = __hmul2(v, {d, d});
|
|
#else
|
|
v.x = (v.x - 8.0f) * d;
|
|
v.y = (v.y - 8.0f) * d;
|
|
#endif // GGML_CUDA_F16
|
|
}
|
|
|
|
static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
|
|
const block_q4_1 * x = (const block_q4_1 *) vx;
|
|
|
|
const dfloat d = __low2half(x[ib].dm);
|
|
const dfloat m = __high2half(x[ib].dm);
|
|
|
|
const int vui = x[ib].qs[iqs];
|
|
|
|
v.x = vui & 0xF;
|
|
v.y = vui >> 4;
|
|
|
|
#ifdef GGML_CUDA_F16
|
|
v = __hmul2(v, {d, d});
|
|
v = __hadd2(v, {m, m});
|
|
#else
|
|
v.x = (v.x * d) + m;
|
|
v.y = (v.y * d) + m;
|
|
#endif // GGML_CUDA_F16
|
|
}
|
|
|
|
static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
|
|
const block_q5_0 * x = (const block_q5_0 *) vx;
|
|
|
|
const dfloat d = x[ib].d;
|
|
|
|
uint32_t qh;
|
|
memcpy(&qh, x[ib].qh, sizeof(qh));
|
|
|
|
const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
|
const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
|
|
|
v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
|
|
v.y = ((x[ib].qs[iqs] >> 4) | xh_1);
|
|
|
|
#ifdef GGML_CUDA_F16
|
|
v = __hsub2(v, {16.0f, 16.0f});
|
|
v = __hmul2(v, {d, d});
|
|
#else
|
|
v.x = (v.x - 16.0f) * d;
|
|
v.y = (v.y - 16.0f) * d;
|
|
#endif // GGML_CUDA_F16
|
|
}
|
|
|
|
static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
|
|
const block_q5_1 * x = (const block_q5_1 *) vx;
|
|
|
|
const dfloat d = __low2half(x[ib].dm);
|
|
const dfloat m = __high2half(x[ib].dm);
|
|
|
|
uint32_t qh;
|
|
memcpy(&qh, x[ib].qh, sizeof(qh));
|
|
|
|
const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
|
const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
|
|
|
v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
|
|
v.y = ((x[ib].qs[iqs] >> 4) | xh_1);
|
|
|
|
#ifdef GGML_CUDA_F16
|
|
v = __hmul2(v, {d, d});
|
|
v = __hadd2(v, {m, m});
|
|
#else
|
|
v.x = (v.x * d) + m;
|
|
v.y = (v.y * d) + m;
|
|
#endif // GGML_CUDA_F16
|
|
}
|
|
|
|
static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
|
|
const block_q8_0 * x = (const block_q8_0 *) vx;
|
|
|
|
const dfloat d = x[ib].d;
|
|
|
|
v.x = x[ib].qs[iqs + 0];
|
|
v.y = x[ib].qs[iqs + 1];
|
|
|
|
#ifdef GGML_CUDA_F16
|
|
v = __hmul2(v, {d, d});
|
|
#else
|
|
v.x *= d;
|
|
v.y *= d;
|
|
#endif // GGML_CUDA_F16
|
|
}
|