mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-11-14 06:49:54 +00:00
Merge branch 'ggerganov:master' into k-shift2
This commit is contained in:
MaggotHATE
GitHub
commit
c616263b01
@ -230,7 +230,7 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
|
||||
|
||||
def parse_args() -> argparse.Namespace:
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
|
||||
description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
|
||||
parser.add_argument(
|
||||
"--outfile", type=Path,
|
||||
help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
|
||||
@ -257,11 +257,11 @@ def parse_args() -> argparse.Namespace:
|
||||
)
|
||||
parser.add_argument(
|
||||
"--base", type=Path, required=True,
|
||||
help="directory containing base model file",
|
||||
help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required",
|
||||
)
|
||||
parser.add_argument(
|
||||
"lora_path", type=Path,
|
||||
help="directory containing LoRA adapter file",
|
||||
help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
|
||||
)
|
||||
|
||||
return parser.parse_args()
|
||||
|
||||
@ -341,6 +341,15 @@ These options help improve the performance and memory usage of the LLaMA models.
|
||||
|
||||
For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
|
||||
|
||||
## LoRA (Low-Rank Adaptation) adapters
|
||||
|
||||
- `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
|
||||
- `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
|
||||
|
||||
You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
|
||||
|
||||
LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
|
||||
|
||||
## Additional Options
|
||||
|
||||
These options provide extra functionality and customization when running the LLaMA models:
|
||||
@ -349,6 +358,4 @@ These options provide extra functionality and customization when running the LLa
|
||||
- `--verbose-prompt`: Print the prompt before generating text.
|
||||
- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
|
||||
- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
|
||||
- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
|
||||
- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
|
||||
- `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable or in an OS-specific local cache.
|
||||
|
||||
@ -991,6 +991,73 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
|
||||
}
|
||||
}
|
||||
return;
|
||||
#elif defined(__riscv_v_intrinsic)
|
||||
if (__riscv_vlenb() >= QK4_0) {
|
||||
const size_t vl = QK4_0;
|
||||
|
||||
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
|
||||
|
||||
vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
for (int l = 0; l < nb; l++) {
|
||||
const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
|
||||
const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
|
||||
const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
|
||||
const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
|
||||
|
||||
const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
|
||||
const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
|
||||
const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
|
||||
const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
|
||||
const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
|
||||
const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
|
||||
const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
|
||||
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
// vector version needs Zvfhmin extension
|
||||
const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
|
||||
const float b_scales[8] = {
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[0]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[1]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[2]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[3]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[4]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[5]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[6]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[7])
|
||||
};
|
||||
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
|
||||
sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
__riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
|
||||
}
|
||||
return;
|
||||
}
|
||||
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
|
||||
{
|
||||
float sumf[8];
|
||||
@ -3171,6 +3238,207 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
|
||||
}
|
||||
}
|
||||
}
|
||||
return;
|
||||
}
|
||||
#elif defined(__riscv_v_intrinsic)
|
||||
if (__riscv_vlenb() >= QK4_0) {
|
||||
const size_t vl = QK4_0;
|
||||
|
||||
for (int y = 0; y < nr / 4; y++) {
|
||||
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
|
||||
vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
for (int l = 0; l < nb; l++) {
|
||||
const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
|
||||
const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
|
||||
const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
|
||||
const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
|
||||
const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
|
||||
const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
|
||||
const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
|
||||
|
||||
// vector version needs Zvfhmin extension
|
||||
const float a_scales[4] = {
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[0]),
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[1]),
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[2]),
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[3])
|
||||
};
|
||||
const float b_scales[8] = {
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[0]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[1]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[2]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[3]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[4]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[5]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[6]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[7])
|
||||
};
|
||||
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
|
||||
|
||||
const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
|
||||
const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
|
||||
const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
|
||||
const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l0;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l0 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
|
||||
sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
|
||||
const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
|
||||
const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
|
||||
const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
|
||||
const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l1;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l1 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
|
||||
sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
|
||||
const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
|
||||
const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
|
||||
const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
|
||||
const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l2;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l2 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
|
||||
sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
|
||||
const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
|
||||
const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
|
||||
const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
|
||||
const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l3;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l3 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
|
||||
sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
}
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
|
||||
}
|
||||
}
|
||||
|
||||
return;
|
||||
}
|
||||
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
|
||||
|
||||
@ -22136,7 +22136,11 @@ static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
|
||||
return false;
|
||||
}
|
||||
|
||||
p->data = GGML_CALLOC(p->n + 1, 1);
|
||||
p->data = calloc(p->n + 1, 1);
|
||||
if (!p->data) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for string of length %" PRIu64 "\n", __func__, p->n);
|
||||
return false;
|
||||
}
|
||||
|
||||
ok = ok && gguf_fread_el(file, p->data, p->n, offset);
|
||||
|
||||
@ -22170,7 +22174,11 @@ static void gguf_free_kv(struct gguf_kv * kv) {
|
||||
}
|
||||
|
||||
struct gguf_context * gguf_init_empty(void) {
|
||||
struct gguf_context * ctx = GGML_CALLOC(1, sizeof(struct gguf_context));
|
||||
struct gguf_context * ctx = calloc(1, sizeof(struct gguf_context));
|
||||
if (!ctx) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
memcpy(ctx->header.magic, GGUF_MAGIC, sizeof(ctx->header.magic));
|
||||
ctx->header.version = GGUF_VERSION;
|
||||
@ -22216,7 +22224,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
|
||||
|
||||
bool ok = true;
|
||||
|
||||
struct gguf_context * ctx = GGML_CALLOC(1, sizeof(struct gguf_context));
|
||||
struct gguf_context * ctx = calloc(1, sizeof(struct gguf_context));
|
||||
if (!ctx) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
|
||||
fclose(file);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
// read the header
|
||||
{
|
||||
@ -22255,9 +22268,13 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
|
||||
{
|
||||
const uint64_t n_kv = ctx->header.n_kv;
|
||||
|
||||
// header.n_kv will hold the actual value of pairs that were successfully read in the loop below
|
||||
ctx->header.n_kv = 0;
|
||||
ctx->kv = GGML_CALLOC(n_kv, sizeof(struct gguf_kv));
|
||||
ctx->kv = calloc(n_kv, sizeof(struct gguf_kv));
|
||||
if (!ctx->kv) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for kv pairs\n", __func__);
|
||||
fclose(file);
|
||||
gguf_free(ctx);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
for (uint64_t i = 0; i < n_kv; ++i) {
|
||||
struct gguf_kv * kv = &ctx->kv[i];
|
||||
@ -22308,7 +22325,13 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
|
||||
return NULL;
|
||||
}
|
||||
|
||||
kv->value.arr.data = GGML_CALLOC(kv->value.arr.n, gguf_type_size(kv->value.arr.type));
|
||||
kv->value.arr.data = calloc(kv->value.arr.n, gguf_type_size(kv->value.arr.type));
|
||||
if (!kv->value.arr.data) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
|
||||
fclose(file);
|
||||
gguf_free(ctx);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
ok = ok && gguf_fread_el(file, kv->value.arr.data, kv->value.arr.n * gguf_type_size(kv->value.arr.type), &offset);
|
||||
} break;
|
||||
@ -22322,24 +22345,36 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
|
||||
return NULL;
|
||||
}
|
||||
|
||||
kv->value.arr.data = GGML_CALLOC(kv->value.arr.n, sizeof(struct gguf_str));
|
||||
kv->value.arr.data = calloc(kv->value.arr.n, sizeof(struct gguf_str));
|
||||
if (!kv->value.arr.data) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
|
||||
fclose(file);
|
||||
gguf_free(ctx);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
|
||||
ok = ok && gguf_fread_str(file, &((struct gguf_str *) kv->value.arr.data)[j], &offset);
|
||||
}
|
||||
} break;
|
||||
case GGUF_TYPE_ARRAY:
|
||||
default: GGML_ABORT("invalid type");
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: invalid array type %d\n", __func__, kv->value.arr.type);
|
||||
ok = false;
|
||||
} break;
|
||||
}
|
||||
} break;
|
||||
default: GGML_ABORT("invalid type");
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: invalid type %d\n", __func__, kv->type);
|
||||
ok = false;
|
||||
} break;
|
||||
}
|
||||
|
||||
if (!ok) {
|
||||
break;
|
||||
}
|
||||
|
||||
ctx->header.n_kv++;
|
||||
}
|
||||
|
||||
if (!ok) {
|
||||
@ -22352,7 +22387,13 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
|
||||
|
||||
// read the tensor infos
|
||||
if (ctx->header.n_tensors > 0) {
|
||||
ctx->infos = GGML_CALLOC(ctx->header.n_tensors, sizeof(struct gguf_tensor_info));
|
||||
ctx->infos = calloc(ctx->header.n_tensors, sizeof(struct gguf_tensor_info));
|
||||
if (!ctx->infos) {
|
||||
fprintf(stderr, "%s: failed to allocate memory for tensor infos\n", __func__);
|
||||
fclose(file);
|
||||
gguf_free(ctx);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
|
||||
struct gguf_tensor_info * info = &ctx->infos[i];
|
||||
|
||||
Loading…
Reference in New Issue
Block a user