chore : correct typos [no ci] (#20041)
* fix(docs): correct typos found during code review Non-functional changes only: - Fixed minor spelling mistakes in comments - Corrected typos in user-facing strings - No variables, logic, or functional code was modified. Signed-off-by: Marcel Petrick <mail@marcelpetrick.it> * Update docs/backend/CANN.md Co-authored-by: Aaron Teo <taronaeo@gmail.com> * Revert "Auxiliary commit to revert individual files from 846d1c301281178efbc6ce6060ad34c1ebe45af8" This reverts commit 02fcf0c7db661d5ff3eff96b2b2db9fdb7213256. * Update tests/test-backend-ops.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update tests/test-backend-ops.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Signed-off-by: Marcel Petrick <mail@marcelpetrick.it> Co-authored-by: Aaron Teo <taronaeo@gmail.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Marcel Petrick
GitHub
parent
7a99dc85e2
commit
92f7da00b4
|
|
@ -159,7 +159,7 @@ Maintainers reserve the right to decline review or close pull requests for any r
|
|||
|
||||
# Code maintenance
|
||||
|
||||
- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file reponsible for:
|
||||
- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file responsible for:
|
||||
- Reviewing and merging related PRs
|
||||
- Fixing related bugs
|
||||
- Providing developer guidance/support
|
||||
|
|
|
|||
|
|
@ -2399,7 +2399,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
|
|||
params.fit_params = false;
|
||||
} else {
|
||||
throw std::runtime_error(
|
||||
string_format("error: unkown value for --fit: '%s'\n", value.c_str()));
|
||||
string_format("error: unknown value for --fit: '%s'\n", value.c_str()));
|
||||
}
|
||||
}
|
||||
).set_env("LLAMA_ARG_FIT"));
|
||||
|
|
|
|||
|
|
@ -869,7 +869,7 @@ std::string common_detokenize(
|
|||
// Embedding utils
|
||||
//
|
||||
|
||||
// TODO: repace embd_norm with an enum
|
||||
// TODO: replace embd_norm with an enum
|
||||
void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
|
||||
|
||||
float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
|
||||
|
|
|
|||
|
|
@ -18,7 +18,7 @@ template <bool abort_on_nan> void common_debug_print_tensor(uint8_t * data, ggml
|
|||
// prints tensors that are processed in the computation graph
|
||||
// by default prints all tensors, but can be configured by creating a `base_callback_data` instance with
|
||||
// non-empty filter_patterns. See examples/debug.ccp for possible usage patterns
|
||||
// The template parameter determins whether an error should be thrown whenever a NaN is encountered
|
||||
// The template parameter determines whether an error should be thrown whenever a NaN is encountered
|
||||
// in a tensor (useful for stopping debug sessions on first erroneous tensor)
|
||||
// The callback data will be passed as the third parameter (user_data)
|
||||
template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);
|
||||
|
|
|
|||
|
|
@ -63,7 +63,7 @@ The llama.cpp Jinja engine introduces `jinja::string` (see `jinja/string.h`), wh
|
|||
- **One-to-many** (e.g., split): result is marked `is_input` **only if ALL** input parts are marked `is_input`
|
||||
- **Many-to-one** (e.g., join): same as one-to-many
|
||||
|
||||
For string concatenation, string parts will be appended to the new string as-is, while perserving the `is_input` flag.
|
||||
For string concatenation, string parts will be appended to the new string as-is, while preserving the `is_input` flag.
|
||||
|
||||
**Enabling Input Marking:**
|
||||
|
||||
|
|
|
|||
|
|
@ -4031,7 +4031,7 @@ class Qwen2VLVisionModel(MmprojModel):
|
|||
# split Conv3D into Conv2Ds
|
||||
c1, c2, kt, kh, kw = data_torch.shape
|
||||
del c1, c2, kh, kw # unused
|
||||
assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
|
||||
assert kt == 2, "Current implementation only support temporal_patch_size of 2"
|
||||
yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight" , data_torch[:, :, 0, ...])
|
||||
yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...])
|
||||
else:
|
||||
|
|
@ -5404,7 +5404,7 @@ class KimiLinearModel(TextModel):
|
|||
# Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv
|
||||
linear_attn_config = self.hparams["linear_attn_config"]
|
||||
# n_head == 0 for KDA layers, n_head > 0 for MLA layers
|
||||
# full_attention_layers list will be used to distingush layer type
|
||||
# full_attention_layers list will be used to distinguish layer type
|
||||
_num_kv_heads = list()
|
||||
_full_attn_layers = linear_attn_config["full_attn_layers"]
|
||||
for il in range(self.hparams["num_hidden_layers"]):
|
||||
|
|
@ -6505,7 +6505,7 @@ class Gemma3VisionModel(MmprojModel):
|
|||
super().set_gguf_parameters()
|
||||
hparams = self.hparams
|
||||
self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GEMMA3)
|
||||
# default values below are taken from HF tranformers code
|
||||
# default values below are taken from HF transformers code
|
||||
self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
|
||||
self.gguf_writer.add_vision_use_gelu(True)
|
||||
# calculate proj_scale_factor (used by tinygemma3 test model)
|
||||
|
|
@ -7097,7 +7097,7 @@ class Rwkv7Model(TextModel):
|
|||
|
||||
if bid == 0 and "time_mix_a" in new_name:
|
||||
# dummy v0/v1/v2 on first layer
|
||||
# easist way to make llama happy
|
||||
# easiest way to make llama happy
|
||||
yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch)
|
||||
|
||||
yield (new_name, data_torch)
|
||||
|
|
@ -9596,7 +9596,7 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
|
|||
# NOTE: Explicitly include hparam prefix prefix for d_model to
|
||||
# disambiguate with top-level head_dim
|
||||
# NOTE 2: If needed for future models, this can be isolated in a method
|
||||
# to separate the prefix setting and teh keys used
|
||||
# to separate the prefix setting and the keys used
|
||||
self.d_model = self.find_hparam([f"{self.hparam_prefixes[0]}_head_dim", "hidden_size", "d_model"])
|
||||
self.n_group = self.find_hparam(["n_groups", "num_groups"])
|
||||
self.d_inner = self.find_hparam(["expand", "num_heads"]) * self.d_model
|
||||
|
|
@ -9743,7 +9743,7 @@ class NemotronHModel(GraniteHybridModel):
|
|||
self.gguf_writer.add_value_length(self.head_dim)
|
||||
|
||||
# Set feed_forward_length
|
||||
# NOTE: This will trigger an override warning. This is preferrable to
|
||||
# NOTE: This will trigger an override warning. This is preferable to
|
||||
# duplicating all the parent logic
|
||||
if not self.is_moe:
|
||||
n_ff = self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"])
|
||||
|
|
|
|||
|
|
@ -20,7 +20,7 @@
|
|||
|
||||
**Llama.cpp + CANN**
|
||||
|
||||
The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly.
|
||||
The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are integrated to CANN Toolkit and kernels to using Ascend NPU directly.
|
||||
|
||||
## News
|
||||
|
||||
|
|
@ -210,7 +210,7 @@ docker run --name llamacpp --device /dev/davinci0 --device /dev/davinci_manager
|
|||
# and install driver.
|
||||
sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
|
||||
```
|
||||
If the following messaage appers, firmware is installed successfully.
|
||||
If the following message appears, firmware is installed successfully.
|
||||
```sh
|
||||
Firmware package installed successfully!
|
||||
```
|
||||
|
|
|
|||
|
|
@ -708,7 +708,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
|
|||
|
||||
- Remove **build** folder or try a clean-build.
|
||||
|
||||
- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
|
||||
- I can **not** see `[ext_oneapi_level_zero:gpu]` after installing the GPU driver on Linux.
|
||||
|
||||
Please double-check with `sudo sycl-ls`.
|
||||
|
||||
|
|
|
|||
|
|
@ -116,7 +116,7 @@ Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920
|
|||
### Windows
|
||||
|
||||
All artifacts are already installed in the `pkg-snapdragon` folder.
|
||||
To run, adapt below instructions to use Powershell scrits in `scripts/snapdragon/windows`.
|
||||
To run, adapt below instructions to use Powershell scripts in `scripts/snapdragon/windows`.
|
||||
|
||||
## How to Run
|
||||
|
||||
|
|
|
|||
|
|
@ -144,7 +144,7 @@ Once the build is complete HTP ops libraries will be installed like this
|
|||
-a---- 1/22/2026 6:01 PM 4139 libggml-htp.cat
|
||||
```
|
||||
|
||||
The .cat file, the signature and proper certicate installation can be verified with
|
||||
The .cat file, the signature and proper certificate installation can be verified with
|
||||
|
||||
```
|
||||
> signtool.exe verify /v /pa .\pkg-snapdragon\lib\libggml-htp.cat
|
||||
|
|
|
|||
|
|
@ -595,7 +595,7 @@ You can verify that KleidiAI is being used by running
|
|||
```bash
|
||||
./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
|
||||
```
|
||||
If KleidiAI is enabled, the ouput will contain a line similar to:
|
||||
If KleidiAI is enabled, the output will contain a line similar to:
|
||||
```
|
||||
load_tensors: CPU_KLEIDIAI model buffer size = 3474.00 MiB
|
||||
```
|
||||
|
|
@ -699,7 +699,7 @@ To read documentation for how to build on Android, [click here](./android.md)
|
|||
|
||||
## WebGPU [In Progress]
|
||||
|
||||
The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
|
||||
The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `bed1a61`.
|
||||
|
||||
In the llama.cpp directory, build with CMake:
|
||||
|
||||
|
|
|
|||
|
|
@ -281,7 +281,7 @@ llama_print_timings: total time = 5990.25 ms / 202 tokens
|
|||
|
||||
Just the same as above.
|
||||
|
||||
**ouput**
|
||||
**output**
|
||||
```sh
|
||||
encode_image_with_clip: image embedding created: 144 tokens
|
||||
|
||||
|
|
@ -305,7 +305,7 @@ llama_print_timings: total time = 15513.95 ms / 412 tokens
|
|||
## Run on Intel(R) Core(TM) Ultra7 115H
|
||||
### operation system
|
||||
Windows11
|
||||
### comiple
|
||||
### compile
|
||||
```sh
|
||||
make -j32
|
||||
```
|
||||
|
|
|
|||
|
|
@ -2,7 +2,7 @@
|
|||
|
||||
This is a utility intended to help debug a model by registering a callback that
|
||||
logs GGML operations and tensor data. It can also store the generated logits or
|
||||
embeddings as well as the prompt and token ids for comparision with the original
|
||||
embeddings as well as the prompt and token ids for comparison with the original
|
||||
model.
|
||||
|
||||
### Usage
|
||||
|
|
|
|||
|
|
@ -43,12 +43,12 @@ Choose one of the following scheduling methods:
|
|||
- `-b`: Batch size
|
||||
|
||||
### Examples
|
||||
#### Dream architechture:
|
||||
#### Dream architecture:
|
||||
```
|
||||
llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual
|
||||
```
|
||||
|
||||
#### LLaDA architechture:
|
||||
#### LLaDA architecture:
|
||||
```
|
||||
llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual
|
||||
```
|
||||
|
|
|
|||
|
|
@ -52,8 +52,8 @@ highlight llama_hl_info guifg=#77ff2f ctermfg=119
|
|||
" n_prefix: number of lines before the cursor location to include in the local prefix
|
||||
" n_suffix: number of lines after the cursor location to include in the local suffix
|
||||
" n_predict: max number of tokens to predict
|
||||
" t_max_prompt_ms: max alloted time for the prompt processing (TODO: not yet supported)
|
||||
" t_max_predict_ms: max alloted time for the prediction
|
||||
" t_max_prompt_ms: max allotted time for the prompt processing (TODO: not yet supported)
|
||||
" t_max_predict_ms: max allotted time for the prediction
|
||||
" show_info: show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
|
||||
" auto_fim: trigger FIM completion automatically on cursor movement
|
||||
" max_line_suffix: do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
|
||||
|
|
|
|||
|
|
@ -69,7 +69,7 @@ Command line arguments take precedence over environment variables when both are
|
|||
|
||||
In cases where the transformer implementation for the model has not been released
|
||||
yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
|
||||
will then cause the transformer implementation to be loaded explicitely and not
|
||||
will then cause the transformer implementation to be loaded explicitly and not
|
||||
use AutoModelForCausalLM:
|
||||
```
|
||||
export UNRELEASED_MODEL_NAME=SomeNewModel
|
||||
|
|
@ -120,7 +120,7 @@ The converted model can be inspected using the following command:
|
|||
(venv) $ make causal-run-converted-model
|
||||
```
|
||||
|
||||
### Model logits verfication
|
||||
### Model logits verification
|
||||
The following target will run the original model and the converted model and
|
||||
compare the logits:
|
||||
```console
|
||||
|
|
@ -235,7 +235,7 @@ new model the model can be converted to GGUF format using the following command:
|
|||
(venv) $ make embedding-run-converted-model
|
||||
```
|
||||
|
||||
### Model logits verfication
|
||||
### Model logits verification
|
||||
The following target will run the original model and the converted model (which
|
||||
was done manually in the previous steps) and compare the logits:
|
||||
```console
|
||||
|
|
@ -335,7 +335,7 @@ $ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LO
|
|||
|
||||
## HuggingFace utilities
|
||||
The following targets are useful for creating collections and model repositories
|
||||
on Hugging Face in the the ggml-org. These can be used when preparing a relase
|
||||
on Hugging Face in the the ggml-org. These can be used when preparing a release
|
||||
to script the process for new model releases.
|
||||
|
||||
For the following targets a `HF_TOKEN` environment variable is required.
|
||||
|
|
@ -347,7 +347,7 @@ For the following targets a `HF_TOKEN` environment variable is required.
|
|||
> $ unset HF_TOKEN
|
||||
|
||||
### Create a new Hugging Face Model (model repository)
|
||||
This will create a new model repsository on Hugging Face with the specified
|
||||
This will create a new model repository on Hugging Face with the specified
|
||||
model name.
|
||||
```console
|
||||
(venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
|
||||
|
|
|
|||
|
|
@ -6,11 +6,11 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
|
|||
|
||||
|Tool Name| Function|Status|
|
||||
|-|-|-|
|
||||
|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
|
||||
|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, etc.|Support|
|
||||
|
||||
### llama-ls-sycl-device
|
||||
|
||||
List all SYCL devices with ID, compute capability, max work group size, ect.
|
||||
List all SYCL devices with ID, compute capability, max work group size, etc.
|
||||
|
||||
1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
|
||||
|
||||
|
|
|
|||
|
|
@ -259,7 +259,7 @@ extern "C" {
|
|||
Example usage:
|
||||
|
||||
// operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
|
||||
// preferrably to run on the same backend as the buffer
|
||||
// preferably to run on the same backend as the buffer
|
||||
ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
|
||||
|
||||
sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);
|
||||
|
|
|
|||
|
|
@ -138,7 +138,7 @@ extern "C" {
|
|||
GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
|
||||
GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
|
||||
|
||||
// set gradients to zero, initilize loss, and optionally reset the optimizer
|
||||
// set gradients to zero, initialize loss, and optionally reset the optimizer
|
||||
GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
|
||||
|
||||
GGML_API bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx); // whether the graphs are allocated_statically
|
||||
|
|
|
|||
|
|
@ -2575,7 +2575,7 @@ extern "C" {
|
|||
struct ggml_tensor * grad,
|
||||
struct ggml_tensor * sgd_params); // alpha, weight decay
|
||||
|
||||
// build forward mutiple tensors and select one of them for computing
|
||||
// build forward multiple tensors and select one of them for computing
|
||||
// this is useful for creating graphs that have constant topology but compute different things based on the input
|
||||
// ref: https://github.com/ggml-org/llama.cpp/pull/18550
|
||||
//
|
||||
|
|
|
|||
|
|
@ -195,7 +195,7 @@ struct tile_config_t{
|
|||
// will be needed.
|
||||
//
|
||||
// Here another commonly used pattern 1-3-3 is skipped, as it is mostly used when m <=16;
|
||||
// and the sinlge batch gemm (m=1) has a special fast path with `avx512-vnni`.
|
||||
// and the single batch gemm (m=1) has a special fast path with `avx512-vnni`.
|
||||
//
|
||||
// ref: https://www.intel.com/content/www/us/en/developer/articles/code-sample/
|
||||
// advanced-matrix-extensions-intrinsics-functions.html
|
||||
|
|
@ -1379,8 +1379,8 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
|
|||
// sum of offsets, shared across COLS
|
||||
//
|
||||
// avx512-vnni does not have `_mm512_dpbssd_epi32`,
|
||||
// need to transfrom ss to us:
|
||||
// a * (b - 8) is equavilent to b * a - 8 * a
|
||||
// need to transform ss to us:
|
||||
// a * (b - 8) is equivalent to b * a - 8 * a
|
||||
// s u u u s u s
|
||||
//
|
||||
__m512i vcomp;
|
||||
|
|
|
|||
|
|
@ -968,7 +968,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
|
|||
|
||||
const int vector_length = ggml_cpu_get_sve_cnt()*8;
|
||||
|
||||
//VLA Implemenation for SVE
|
||||
//VLA Implementation for SVE
|
||||
switch (vector_length) {
|
||||
case 128:
|
||||
{
|
||||
|
|
|
|||
|
|
@ -781,7 +781,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int n,
|
|||
|
||||
const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
|
||||
|
||||
// Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
|
||||
// Load the 64 quants from q8K duplicated to use vecdots with the interleaved columns
|
||||
// but still need the qs to use the low and hi bits from q4
|
||||
const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
|
||||
int8x16_t q8_qs[8];
|
||||
|
|
@ -3796,7 +3796,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n,
|
|||
|
||||
for (int b = 0; b < nb; b++) {
|
||||
// bsums pairs belongs to the same q8_k subblock
|
||||
// 64 elemnts loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
|
||||
// 64 elements loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
|
||||
const int16x8_t bsums[4]{
|
||||
vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
|
||||
vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
|
||||
|
|
|
|||
|
|
@ -423,7 +423,7 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
|
|||
quants_interleaved[j] = i0;
|
||||
}
|
||||
|
||||
// Masks to shuffle the quants of corresonding sub blocks for rearraning quants for vectorized bsums computation
|
||||
// Masks to shuffle the quants of corresponding sub blocks for rearranging quants for vectorized bsums computation
|
||||
__m256i shuffle_mask_sb2 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 0, 1, 4, 5, 6, 7, 8, 9, 8, 9, 12, 13, 14, 15));
|
||||
shuffle_mask_sb2 = _mm256_permute2f128_si256(shuffle_mask_sb2, shuffle_mask_sb2, 0);
|
||||
__m256i shuffle_mask_sb3 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 0, 1, 6, 7, 8, 9, 10, 11, 8, 9, 14, 15));
|
||||
|
|
@ -625,7 +625,7 @@ static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
|||
iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_3 ,_mm256_shuffle_epi32(rhs_vec_4567_3, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170));
|
||||
iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255));
|
||||
|
||||
// Accumulated values multipled with appropriate scales
|
||||
// Accumulated values multiplied with appropriate scales
|
||||
acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
|
||||
}
|
||||
|
||||
|
|
@ -868,7 +868,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
|||
const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
|
||||
const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
|
|
@ -1076,7 +1076,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
|||
const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
|
||||
const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
|
|
@ -1257,7 +1257,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
|||
// Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
|
||||
const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
|
|
@ -1428,7 +1428,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
|
|||
// Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
|
||||
const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate
|
||||
// Multiply with appropriate scales and accumulate
|
||||
acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
|
|
@ -1612,7 +1612,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
|
||||
|
||||
// Dot product done within 32 bit lanes and accumulated in the same vector
|
||||
// First done for first sub block and thenn for second sub block in each sb
|
||||
// First done for first sub block and then for second sub block in each sb
|
||||
// B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
|
||||
// B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
|
||||
// ...........................................................................
|
||||
|
|
@ -2422,7 +2422,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
|
|
@ -2785,7 +2785,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
|
|
@ -2802,7 +2802,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
|
||||
}
|
||||
}
|
||||
// Store accumlated values
|
||||
// Store accumulated values
|
||||
for (int i = 0; i < 4; i++) {
|
||||
_mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
|
||||
}
|
||||
|
|
@ -3130,7 +3130,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);//GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
|
|
@ -3460,7 +3460,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); //GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
|
|
@ -4268,7 +4268,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
|
|
@ -5035,7 +5035,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
|
||||
}
|
||||
}
|
||||
// Store accumlated values
|
||||
// Store accumulated values
|
||||
for (int i = 0; i < 4; i++) {
|
||||
_mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
|
||||
}
|
||||
|
|
@ -5677,7 +5677,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
|
||||
acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
|
||||
acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
|
||||
|
|
@ -6349,7 +6349,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
|||
const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
|
||||
const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
|
||||
|
||||
// Multiply with appropiate scales and accumulate (for both d and dmin) below
|
||||
// Multiply with appropriate scales and accumulate (for both d and dmin) below
|
||||
acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
|
||||
acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
|
||||
acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
|
||||
|
|
|
|||
|
|
@ -2477,7 +2477,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
|
|||
|
||||
if (prio != GGML_SCHED_PRIO_LOW) {
|
||||
// Tell Windows that this thread should not be throttled (needs its own CPU core).
|
||||
// Newer Windows 11 versions aggresively park (offline) CPU cores and often place
|
||||
// Newer Windows 11 versions aggressively park (offline) CPU cores and often place
|
||||
// all our threads onto the first 4 cores which results in terrible performance with
|
||||
// n_threads > 4
|
||||
#if _WIN32_WINNT >= 0x0602
|
||||
|
|
|
|||
|
|
@ -533,7 +533,7 @@ class tinyBLAS {
|
|||
if constexpr (RN > 1) {
|
||||
return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
|
||||
} else {
|
||||
GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
|
||||
GGML_LOG_ERROR("mnpack<%d, %d> block size not supported\n", RM, (int)SIZE_N);
|
||||
GGML_ASSERT(false); // we have miss something.
|
||||
}
|
||||
}
|
||||
|
|
@ -711,7 +711,7 @@ class tinyBLAS_RVV {
|
|||
if constexpr (RN > 1) {
|
||||
return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
|
||||
} else {
|
||||
GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
|
||||
GGML_LOG_ERROR("mnpack<%d, %d> block size not supported\n", RM, (int)SIZE_N);
|
||||
GGML_ASSERT(false); // we have miss something.
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -375,7 +375,7 @@ static void ggml_compute_forward_dup_bytes(
|
|||
const size_t rs = ne00 * type_size;
|
||||
|
||||
if (nb00 == type_size) {
|
||||
// src0 is contigous on first dimension, copy by rows
|
||||
// src0 is contiguous on first dimension, copy by rows
|
||||
for (int64_t i03 = 0; i03 < ne03; i03++) {
|
||||
for (int64_t i02 = 0; i02 < ne02; i02++) {
|
||||
id += rs * ir0;
|
||||
|
|
@ -1795,7 +1795,7 @@ void ggml_compute_forward_repeat(
|
|||
{
|
||||
ggml_compute_forward_repeat_f32(params, dst);
|
||||
} break;
|
||||
// TODO: templateify the implemenation and support for I64
|
||||
// TODO: templateify the implementation and support for I64
|
||||
// ref https://github.com/ggml-org/llama.cpp/pull/14274#discussion_r2169492225
|
||||
//case GGML_TYPE_I64:
|
||||
// {
|
||||
|
|
|
|||
|
|
@ -3032,7 +3032,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
|
|||
case GGML_OP_MUL_MAT_ID:
|
||||
{
|
||||
size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
|
||||
size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
|
||||
size = GGML_PAD(size, sizeof(int64_t)); // + padding for next block.
|
||||
|
||||
const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
|
||||
const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
|
||||
|
|
@ -3297,7 +3297,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
|
|||
auto * wdata = (char *)params->wdata;
|
||||
auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));
|
||||
|
||||
// total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t)
|
||||
// total of [n_as][ne12 + 1] elements of type mmid_row_mapping (2*int32_t = int64_t)
|
||||
auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
|
||||
struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
|
||||
|
||||
|
|
|
|||
|
|
@ -1215,7 +1215,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
|
|||
}
|
||||
|
||||
// If attention sinks are used, potentially re-scale if KQ_max is small.
|
||||
// Also add the sink as a value to KQ_rowsum, this is done after synchonization of KQ_rowsum
|
||||
// Also add the sink as a value to KQ_rowsum, this is done after synchronization of KQ_rowsum
|
||||
// so it's being done unconditionally for every thread.
|
||||
if (!is_fixup && (np == 1 || threadIdx.y % np == 0) && sinks_f) {
|
||||
float KQ_max_scale[cols_per_thread];
|
||||
|
|
|
|||
|
|
@ -10,7 +10,7 @@ static constexpr __device__ int ggml_cuda_fattn_vec_get_nthreads_device() {
|
|||
return 128;
|
||||
}
|
||||
|
||||
// Currenlty llvm with the amdgcn target does not support unrolling loops
|
||||
// Currently llvm with the amdgcn target does not support unrolling loops
|
||||
// that contain a break that can not be resolved at compile time.
|
||||
#ifdef __clang__
|
||||
#pragma clang diagnostic push
|
||||
|
|
|
|||
|
|
@ -18,7 +18,7 @@
|
|||
#if defined(RDNA4) && ROCWMMA_VERSION_MAJOR > 1
|
||||
#define GGML_USE_WMMA_FATTN
|
||||
#elif defined(RDNA4)
|
||||
#warning "rocwmma fattn is not suported on RDNA4 on rocwmma < v2.0.0, expect degraded performance"
|
||||
#warning "rocwmma fattn is not supported on RDNA4 on rocwmma < v2.0.0, expect degraded performance"
|
||||
#endif // defined(RDNA4) && ROCWMMA_VERSION_MAJOR > 1
|
||||
#endif // defined(GGML_HIP_ROCWMMA_FATTN)
|
||||
|
||||
|
|
|
|||
|
|
@ -3330,7 +3330,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph,
|
|||
return false;
|
||||
}
|
||||
|
||||
//rms_norm kernel assumes contigous rows
|
||||
//rms_norm kernel assumes contiguous rows
|
||||
if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
|
||||
return false;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -235,7 +235,7 @@ static __global__ void quantize_mmq_q8_1(
|
|||
q.z = roundf(xi.z*d_inv);
|
||||
q.w = roundf(xi.w*d_inv);
|
||||
|
||||
// Write back 4 int8 values as a single 32 bit value for better memroy bandwidth:
|
||||
// Write back 4 int8 values as a single 32 bit value for better memory bandwidth:
|
||||
char4 * yqs4 = (char4 *) y[ib].qs;
|
||||
yqs4[iqs/4] = q;
|
||||
|
||||
|
|
|
|||
|
|
@ -46,7 +46,7 @@ struct soft_max_params {
|
|||
};
|
||||
|
||||
// When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
|
||||
// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
|
||||
// As we want to keep pragma unroll for all other cases we suppress the clang transformation warning here.
|
||||
#ifdef __clang__
|
||||
#pragma clang diagnostic push
|
||||
#pragma clang diagnostic ignored "-Wpass-failed"
|
||||
|
|
|
|||
|
|
@ -83,7 +83,7 @@ static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
|
|||
// ======================
|
||||
// When ncols_template == 0 the bounds for the loops in this function are not
|
||||
// known and can't be unrolled. As we want to keep pragma unroll for all other
|
||||
// cases we supress the clang transformation warning here.
|
||||
// cases we suppress the clang transformation warning here.
|
||||
#ifdef __clang__
|
||||
# pragma clang diagnostic push
|
||||
# pragma clang diagnostic ignored "-Wpass-failed"
|
||||
|
|
|
|||
|
|
@ -139,7 +139,7 @@ struct ggml_hexagon_session {
|
|||
};
|
||||
|
||||
void ggml_hexagon_session::enqueue(struct htp_general_req &req, struct dspqueue_buffer *bufs, uint32_t n_bufs, bool sync) {
|
||||
// Bump pending flag (cleared in the session::flush once we get the responce)
|
||||
// Bump pending flag (cleared in the session::flush once we get the response)
|
||||
this->op_pending++; // atomic inc
|
||||
|
||||
int err = dspqueue_write(this->queue,
|
||||
|
|
@ -443,7 +443,7 @@ static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
|
|||
|
||||
// Repack the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Repack the scales
|
||||
ggml_half * d = (ggml_half *) (y_d + i * dblk_size);
|
||||
|
|
@ -503,7 +503,7 @@ static void unpack_row_q4x4x2(block_q4_0 * x, const uint8_t * y, int64_t k) {
|
|||
|
||||
// Repack the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Unpack the scales
|
||||
const ggml_half * d = (const ggml_half *) (y_d + i * dblk_size);
|
||||
|
|
@ -552,7 +552,7 @@ static void init_row_q4x4x2(block_q4_0 * x, int64_t k) {
|
|||
|
||||
// Init the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Unpack the scales
|
||||
x[i * 8 + 0].d = 0;
|
||||
|
|
@ -770,7 +770,7 @@ static void repack_row_q8x4x2(uint8_t * y, const block_q8_0 * x, int64_t k) {
|
|||
|
||||
// Repack the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Repack the scales
|
||||
ggml_half * d = (ggml_half *) (y_d + i * dblk_size);
|
||||
|
|
@ -829,7 +829,7 @@ static void unpack_row_q8x4x2(block_q8_0 * x, const uint8_t * y, int64_t k) {
|
|||
|
||||
// Repack the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q4_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Unpack the scales
|
||||
const ggml_half * d = (const ggml_half *) (y_d + i * dblk_size);
|
||||
|
|
@ -878,7 +878,7 @@ static void init_row_q8x4x2(block_q8_0 * x, int64_t k) {
|
|||
|
||||
// Init the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_Q8_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Unpack the scales
|
||||
x[i * 8 + 0].d = 0;
|
||||
|
|
@ -1120,7 +1120,7 @@ static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k)
|
|||
|
||||
// Repack the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_MXFP4x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Repack the scales
|
||||
uint8_t * e = (uint8_t *) (y_e + i * eblk_size);
|
||||
|
|
@ -1180,7 +1180,7 @@ static void unpack_row_mxfp4x4x2(block_mxfp4 * x, const uint8_t * y, int64_t k)
|
|||
|
||||
// Repack the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_MXFP4_0x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Unpack the scales
|
||||
const uint8_t * e = (const uint8_t *) (y_e + i * eblk_size);
|
||||
|
|
@ -1229,7 +1229,7 @@ static void init_row_mxfp4x4x2(block_mxfp4 * x, int64_t k) {
|
|||
|
||||
// Init the scales
|
||||
// Note: Do not combine with the loop above. For tensor sizes not multiple of 256 (QK_MXFP4x4x2)
|
||||
// the last block is truncated and overriden by the scales.
|
||||
// the last block is truncated and overridden by the scales.
|
||||
for (int i = 0; i < nb; i++) {
|
||||
// Unpack the scales
|
||||
x[i * 8 + 0].e = 0;
|
||||
|
|
@ -2670,7 +2670,7 @@ static std::vector<int> ggml_hexagon_graph_optimize_reorder(const std::vector<no
|
|||
// The main goal here is to stack the MUL_MAT ops with the same src1 input.
|
||||
// This allows use to reuse dynamically quantized src1 in VTCM.
|
||||
|
||||
// TODO: the current version might do incorrect reodering in cases where quantized src0
|
||||
// TODO: the current version might do incorrect reordering in cases where quantized src0
|
||||
// input is an output of another Op.
|
||||
|
||||
for (int i0 = 0; i0 < n; i0++) {
|
||||
|
|
|
|||
|
|
@ -282,7 +282,7 @@ static std::string get_driver_path() {
|
|||
// Replace \SystemRoot with an absolute path from system ENV windir
|
||||
const std::wstring systemRootEnv = L"windir";
|
||||
|
||||
// Query the number of wide charactors this variable requires
|
||||
// Query the number of wide characters this variable requires
|
||||
DWORD numWords = GetEnvironmentVariableW(systemRootEnv.c_str(), NULL, 0);
|
||||
if (numWords == 0) {
|
||||
GGML_LOG_ERROR("ggml-hex: Failed get systemRoot environment variable\n");
|
||||
|
|
|
|||
|
|
@ -67,7 +67,7 @@ static inline HVX_Vector hvx_vec_inverse_f16(HVX_Vector vals) {
|
|||
|
||||
HVX_Vector vcl0 = Q6_Vuh_vcl0_Vuh(rm); //count leading zeros
|
||||
|
||||
// Get mantissa for 16-bit represenation
|
||||
// Get mantissa for 16-bit representation
|
||||
HVX_Vector mant_recip = Q6_V_vand_VV(Q6_Vh_vasr_VhR(Q6_Vh_vasl_VhVh(rm, vcl0), 5), Q6_Vh_vsplat_R(0x03FF));
|
||||
|
||||
//Compute Reciprocal Exponent
|
||||
|
|
|
|||
|
|
@ -18,7 +18,7 @@
|
|||
#include "htp-msg.h"
|
||||
#include "htp-ops.h"
|
||||
|
||||
// Redefined the types GGML_ROPE_TYPE_NORMAL & GGML_ROPE_TYPE_NEOX as we cant include ggml.h
|
||||
// Redefined the types GGML_ROPE_TYPE_NORMAL & GGML_ROPE_TYPE_NEOX as we can't include ggml.h
|
||||
#define HTP_ROPE_TYPE_NORMAL 0
|
||||
#define HTP_ROPE_TYPE_NEOX 2
|
||||
|
||||
|
|
|
|||
|
|
@ -56,7 +56,7 @@ static void worker_pool_main(void * context) {
|
|||
unsigned int n = atomic_load(&pool->n_jobs);
|
||||
unsigned int i = atomic_fetch_add(&pool->next_job, 1);
|
||||
if (i >= n) {
|
||||
// Spurios wakeup
|
||||
// Spurious wakeup
|
||||
continue;
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -1281,7 +1281,7 @@ struct ggml_metal_buffer {
|
|||
bool use_residency_sets;
|
||||
|
||||
// optional MTLResidencySet
|
||||
// note: cannot use explicity "id<MTLResidencySet>" here because it is not available on certain OSes
|
||||
// note: cannot use explicitly "id<MTLResidencySet>" here because it is not available on certain OSes
|
||||
id rset;
|
||||
|
||||
// pointers to global device
|
||||
|
|
|
|||
|
|
@ -631,7 +631,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
|
|||
const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
|
||||
|
||||
if (!inplace) {
|
||||
// run a separete kernel to cpy src->dst
|
||||
// run a separate kernel to cpy src->dst
|
||||
// not sure how to avoid this
|
||||
// TODO: make a simpler cpy_bytes kernel
|
||||
|
||||
|
|
@ -1644,7 +1644,7 @@ int ggml_metal_op_set(ggml_metal_op_t ctx, int idx) {
|
|||
const bool inplace = (bool) ((const int32_t *) op->op_params)[4];
|
||||
|
||||
if (!inplace) {
|
||||
// run a separete kernel to cpy src->dst
|
||||
// run a separate kernel to cpy src->dst
|
||||
// not sure how to avoid this
|
||||
// TODO: make a simpler cpy_bytes kernel
|
||||
|
||||
|
|
@ -2005,7 +2005,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
|
|||
const int16_t r0ptg = nypsg*nsg; // num src0 rows per threadgroup
|
||||
int16_t r1ptg = 4; // num src1 rows per threadgroup
|
||||
|
||||
// note: not sure how optimal are those across all different hardware. there might be someting cleverer
|
||||
// note: not sure how optimal are those across all different hardware. there might be something cleverer
|
||||
switch (ne11) {
|
||||
case 2:
|
||||
r1ptg = 2; break;
|
||||
|
|
|
|||
|
|
@ -14,7 +14,7 @@
|
|||
#define GGML_METAL_MAX_DEVICES 16
|
||||
|
||||
// number of Metal devices
|
||||
// note: can be overriden with GGML_METAL_DEVICES env to simulate virtual devices
|
||||
// note: can be overridden with GGML_METAL_DEVICES env to simulate virtual devices
|
||||
static int g_devices = 1;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
|
|
|||
|
|
@ -4218,7 +4218,7 @@ kernel void kernel_im2col(
|
|||
template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
|
||||
template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
|
||||
|
||||
// TODO: obolete -- remove
|
||||
// TODO: obsolete -- remove
|
||||
//typedef void (im2col_ext_t)(
|
||||
// constant ggml_metal_kargs_im2col & args,
|
||||
// device const float * x,
|
||||
|
|
|
|||
|
|
@ -313,7 +313,7 @@ struct ProfilingInfo {
|
|||
cl_ulong cmd_duration_ns;
|
||||
// The time for the kernel to complete - COMPLETE - END
|
||||
cl_ulong cmd_complete_duration_ns;
|
||||
// Total time to finish the kernel - COMPELTE - QUEUED
|
||||
// Total time to finish the kernel - COMPLETE - QUEUED
|
||||
cl_ulong cmd_total_duration_ns;
|
||||
// Global and local work sizes.
|
||||
size_t global_size[3];
|
||||
|
|
@ -2555,7 +2555,7 @@ static std::vector<ggml_backend_device> ggml_opencl_probe_devices(ggml_backend_r
|
|||
|
||||
cl_platform_id platform_ids[NPLAT];
|
||||
if (clGetPlatformIDs(NPLAT, platform_ids, &n_platforms) != CL_SUCCESS) {
|
||||
GGML_LOG_ERROR("ggml_opencl: plaform IDs not available.\n");
|
||||
GGML_LOG_ERROR("ggml_opencl: platform IDs not available.\n");
|
||||
return found_devices;
|
||||
}
|
||||
|
||||
|
|
@ -3339,7 +3339,7 @@ static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
|
|||
CL_CHECK(clReleaseEvent(evt));
|
||||
}
|
||||
|
||||
// Syncronizes the 'backend_ctx's device with others so that commands
|
||||
// Synchronizes the 'backend_ctx's device with others so that commands
|
||||
// enqueued to it won't start until commands in the other devices have
|
||||
// completed.
|
||||
static void sync_with_other_backends(ggml_backend_opencl_context * backend_ctx) {
|
||||
|
|
@ -3997,7 +3997,7 @@ struct ggml_backend_opencl_buffer_context {
|
|||
|
||||
// The buffer_context is initially created by ggml_backend_buft_alloc_buffer
|
||||
// before any tensor is initialized (at the beginning of alloc_tensor_range).
|
||||
// Hence, there is alway a buffer object in this vector. When each tensor is
|
||||
// Hence, there is always a buffer object in this vector. When each tensor is
|
||||
// being initialized, this original buffer object will be released if both
|
||||
// flattening and small allocation are enabled, and additional buffer
|
||||
// objects will be created in init_tensor to represent flattened quantized
|
||||
|
|
@ -4132,7 +4132,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
|
|||
//GGML_ASSERT(offset == 0);
|
||||
|
||||
// We create subbuffers from the original tensor buffer for scales and
|
||||
// quants - i.e., scales and quants are aliases into the buffer obejct
|
||||
// quants - i.e., scales and quants are aliases into the buffer object
|
||||
// that backs the original tensor. This is a cleaner way to adapt to the
|
||||
// new memory management.
|
||||
// In the old code, we allocate new buffers for scales and quants
|
||||
|
|
|
|||
|
|
@ -76,10 +76,10 @@ extern int g_ggml_sycl_prioritize_dmmv;
|
|||
|
||||
|
||||
#define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP
|
||||
#define VER_4VEC 610 // todo for hardward optimize.
|
||||
#define VER_GEN9 700 // todo for hardward optimize.
|
||||
#define VER_GEN12 1000000 // todo for hardward optimize.
|
||||
#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardward optimize.
|
||||
#define VER_4VEC 610 // todo for hardware optimize.
|
||||
#define VER_GEN9 700 // todo for hardware optimize.
|
||||
#define VER_GEN12 1000000 // todo for hardware optimize.
|
||||
#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardware optimize.
|
||||
|
||||
#define GGML_SYCL_MAX_NODES 8192 // TODO: adapt to hardwares
|
||||
|
||||
|
|
|
|||
|
|
@ -29,7 +29,7 @@ namespace ggml_sycl_reordered {
|
|||
// [qs0, qs1, qs2, ..., qsN] [d0, d1, d2, ..., dN]
|
||||
//
|
||||
// Notes: out-of-bounds qs will run into d values
|
||||
// Aligment relies on the allocated size of qs
|
||||
// Alignment relies on the allocated size of qs
|
||||
|
||||
template <ggml_type type> struct block_q_t;
|
||||
|
||||
|
|
|
|||
|
|
@ -37,7 +37,7 @@ struct soft_max_params {
|
|||
};
|
||||
|
||||
// When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
|
||||
// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
|
||||
// As we want to keep pragma unroll for all other cases we suppress the clang transformation warning here.
|
||||
#ifdef __clang__
|
||||
#pragma clang diagnostic push
|
||||
#pragma clang diagnostic ignored "-Wpass-failed"
|
||||
|
|
|
|||
|
|
@ -90,7 +90,7 @@ if (Vulkan_FOUND)
|
|||
target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
|
||||
|
||||
# Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
|
||||
# Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
|
||||
# Possibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
|
||||
if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
|
||||
add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
|
||||
endif()
|
||||
|
|
|
|||
|
|
@ -186,7 +186,7 @@ class Metadata:
|
|||
# Quick hack to fix the Norway problem
|
||||
# https://hitchdev.com/strictyaml/why/implicit-typing-removed/
|
||||
yaml_content = yaml_content.replace("- no\n", "- \"no\"\n")
|
||||
# yaml should use 2 spaces insted of tab
|
||||
# yaml should use 2 spaces instead of tab
|
||||
# this issue has came up with the Qwen/Qwen3-235B-A22B-Instruct-2507 model card
|
||||
# (I've also sent a pr tp fix the modelcard too)
|
||||
yaml_content = yaml_content.replace("\t", " ")
|
||||
|
|
|
|||
|
|
@ -164,7 +164,7 @@ class TestMetadataMethod(unittest.TestCase):
|
|||
self.assertEqual(gguf.Metadata.get_model_id_components("Llama-3-Instruct-abliteration-LoRA-8B"),
|
||||
('Llama-3-Instruct-abliteration-LoRA-8B', None, 'Llama-3', 'Instruct-abliteration-LoRA', None, '8B'))
|
||||
|
||||
# Negative size --> output is a LoRA adaper --> prune "LoRA" out of the name to avoid redundancy with the suffix
|
||||
# Negative size --> output is a LoRA adapter --> prune "LoRA" out of the name to avoid redundancy with the suffix
|
||||
self.assertEqual(gguf.Metadata.get_model_id_components("Llama-3-Instruct-abliteration-LoRA-8B", -1234),
|
||||
('Llama-3-Instruct-abliteration-LoRA-8B', None, 'Llama-3', 'Instruct-abliteration', None, '8B'))
|
||||
|
||||
|
|
|
|||
|
|
@ -973,7 +973,7 @@ extern "C" {
|
|||
|
||||
// Logits for the ith token. For positive indices, Equivalent to:
|
||||
// llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
|
||||
// Negative indicies can be used to access logits in reverse order, -1 is the last logit.
|
||||
// Negative indices can be used to access logits in reverse order, -1 is the last logit.
|
||||
// returns NULL for invalid ids.
|
||||
LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
|
||||
|
||||
|
|
@ -988,7 +988,7 @@ extern "C" {
|
|||
|
||||
// Get the embeddings for the ith token. For positive indices, Equivalent to:
|
||||
// llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
|
||||
// Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding.
|
||||
// Negative indices can be used to access embeddings in reverse order, -1 is the last embedding.
|
||||
// shape: [n_embd] (1-dimensional)
|
||||
// returns NULL for invalid ids.
|
||||
LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
|
||||
|
|
@ -1008,9 +1008,9 @@ extern "C" {
|
|||
// Returns LLAMA_TOKEN_NULL if no token was sampled.
|
||||
LLAMA_API llama_token llama_get_sampled_token_ith(struct llama_context * ctx, int32_t i);
|
||||
|
||||
// Get the backend sampled probabilites for the ith token
|
||||
// Get the backend sampled probabilities for the ith token
|
||||
// The index matches llama_get_sampled_token_ith().
|
||||
// Returns NULL if no probabilites were generated.
|
||||
// Returns NULL if no probabilities were generated.
|
||||
LLAMA_API float * llama_get_sampled_probs_ith (struct llama_context * ctx, int32_t i);
|
||||
LLAMA_API uint32_t llama_get_sampled_probs_count_ith(struct llama_context * ctx, int32_t i);
|
||||
|
||||
|
|
@ -1337,7 +1337,7 @@ extern "C" {
|
|||
float tau,
|
||||
float eta);
|
||||
|
||||
/// @details Intializes a GBNF grammar, see grammars/README.md for details.
|
||||
/// @details Initializes a GBNF grammar, see grammars/README.md for details.
|
||||
/// @param vocab The vocabulary that this grammar will be used with.
|
||||
/// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
|
||||
/// @param grammar_root The name of the start symbol for the grammar.
|
||||
|
|
|
|||
|
|
@ -1,6 +1,6 @@
|
|||
#!/usr/bin/env bash
|
||||
|
||||
# intialize a new worktree from a PR number:
|
||||
# initialize a new worktree from a PR number:
|
||||
#
|
||||
# - creates a new remote using the fork's clone URL
|
||||
# - creates a local branch tracking the remote branch
|
||||
|
|
|
|||
|
|
@ -292,6 +292,6 @@ if __name__ == "__main__":
|
|||
"--n_predict_min", type=int, default=1024,
|
||||
help="Min. number of tokens to predict per prompt (supported for synthetic prompts only)")
|
||||
parser.add_argument("--seed_offset", type=int, default=0, help="Offset for determining the seeds for pseudorandom prompt/generation lengths. "
|
||||
"Corelations between seeds can occur when set >= 1000. Negative values mean no seed.")
|
||||
"Correlations between seeds can occur when set >= 1000. Negative values mean no seed.")
|
||||
args = parser.parse_args()
|
||||
benchmark(**vars(args))
|
||||
|
|
|
|||
|
|
@ -158,7 +158,7 @@ llama_context::llama_context(
|
|||
cparams.op_offload = params.op_offload;
|
||||
cparams.kv_unified = params.kv_unified;
|
||||
|
||||
// intialized later
|
||||
// initialized later
|
||||
cparams.pipeline_parallel = false;
|
||||
|
||||
{
|
||||
|
|
@ -1981,7 +1981,7 @@ ggml_cgraph * llama_context::graph_reserve(
|
|||
|
||||
ggml_backend_sched_reset(sched.get());
|
||||
|
||||
// when the scheduler is reset, we cannnot reuse the old graph, so we reset the previous graph result to prevent that
|
||||
// when the scheduler is reset, we cannot reuse the old graph, so we reset the previous graph result to prevent that
|
||||
gf_res_prev->reset();
|
||||
|
||||
// store the n_outputs as it is, and restore it afterwards
|
||||
|
|
|
|||
|
|
@ -1616,7 +1616,7 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
|
|||
ggml_tensor * llm_graph_context::build_inp_out_ids() const {
|
||||
// note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls,
|
||||
// but this would make the graph topology depend on the number of output tokens, which can interere with
|
||||
// features that require constant topology such as pipline parallelism
|
||||
// features that require constant topology such as pipeline parallelism
|
||||
// ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471
|
||||
//if (n_outputs < n_tokens) {
|
||||
// return nullptr;
|
||||
|
|
@ -1779,7 +1779,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
|
|||
if (v_mla) {
|
||||
#if 0
|
||||
// v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
|
||||
// However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
|
||||
// However, the code is optimized for dimensions 0 and 1 being large, so this is inefficient.
|
||||
cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
|
||||
cur = ggml_mul_mat(ctx0, v_mla, cur);
|
||||
#else
|
||||
|
|
|
|||
|
|
@ -583,7 +583,7 @@ llama_kv_cache::slot_info_vec_t llama_kv_cache::prepare(const std::vector<llama_
|
|||
break;
|
||||
}
|
||||
|
||||
// remeber the position that we found
|
||||
// remember the position that we found
|
||||
res.push_back(sinfo_new);
|
||||
|
||||
// store the old state of the cells in the recovery stack
|
||||
|
|
@ -1293,7 +1293,7 @@ static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float *
|
|||
}
|
||||
|
||||
for (uint32_t s = 0; s < n_stream; ++s) {
|
||||
// bookeeping of the KQ mask cells that could change for other tokens of the same sequence
|
||||
// bookkeeping of the KQ mask cells that could change for other tokens of the same sequence
|
||||
std::unordered_map<llama_seq_id, uint32_t> seq_srct;
|
||||
std::unordered_map<llama_seq_id, std::vector<uint32_t>> seq_idxs;
|
||||
|
||||
|
|
|
|||
|
|
@ -1524,7 +1524,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
|||
}
|
||||
|
||||
switch (hparams.n_layer) {
|
||||
// TODO: Jamba layers are a bit heterogenous, so naming this is hard.
|
||||
// TODO: Jamba layers are a bit heterogeneous, so naming this is hard.
|
||||
case 12: // 900M 8x???M
|
||||
case 32: // 51B 16x?B
|
||||
default: type = LLM_TYPE_UNKNOWN;
|
||||
|
|
|
|||
|
|
@ -1833,7 +1833,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|
|||
const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
|
||||
precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
|
||||
#if defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
|
||||
// correct endiannes of data in precompiled_charsmap binary blob
|
||||
// correct endianness of data in precompiled_charsmap binary blob
|
||||
uint32_t * xcda_blob_size = (uint32_t *) &precompiled_charsmap[0];
|
||||
*xcda_blob_size = __builtin_bswap32(*xcda_blob_size);
|
||||
assert(*xcda_blob_size + sizeof(uint32_t) < n_precompiled_charsmap);
|
||||
|
|
|
|||
|
|
@ -146,7 +146,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
|
|||
cb(Qcur, "Qcur_attn_temp_scaled", il);
|
||||
}
|
||||
|
||||
// note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
|
||||
// note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
|
||||
cur = build_attn(inp_attn_k,
|
||||
model.layers[il].wo, NULL,
|
||||
Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
|
||||
|
|
|
|||
|
|
@ -3,7 +3,7 @@
|
|||
#include "llama-model.h"
|
||||
#include "llama-graph.h"
|
||||
|
||||
// note: almost all graphs require atleast sqrtf, so include cmath globally
|
||||
// note: almost all graphs require at least sqrtf, so include cmath globally
|
||||
#include <cmath>
|
||||
|
||||
//
|
||||
|
|
|
|||
|
|
@ -773,7 +773,7 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
|
|||
// tiny_aya digit grouping pattern from tokenizer.json:
|
||||
// {"type": "Split", "pattern": {"Regex": "\\d{1,3}(?=(?:\\d{3})*\\b)"}, "behavior": "Isolated"}
|
||||
// Splits digits into groups of 3 from the right (e.g., 1234567 -> 1, 234, 567)
|
||||
// TODO: Revisit this regex, incase there are any subtle tokenization differences with the original regex.
|
||||
// TODO: Revisit this regex, in case there are any subtle tokenization differences with the original regex.
|
||||
bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -285,7 +285,7 @@ static void test_max_size_too_many_tensors() {
|
|||
GGML_ASSERT(backend.context->allocated_total() <= 16 + 16);
|
||||
}
|
||||
|
||||
// Scenario where there is some space left in the first buffer, but not enough to accomodate
|
||||
// Scenario where there is some space left in the first buffer, but not enough to accommodate
|
||||
// a larger tensor, so a second buffer is required
|
||||
static void test_max_size_tensor_too_large() {
|
||||
dummy_backend backend = dummy_backend_init(32);
|
||||
|
|
|
|||
|
|
@ -1868,9 +1868,9 @@ struct test_case {
|
|||
};
|
||||
|
||||
|
||||
// ###################################
|
||||
// ## Section 2: GGML Op Defintions ##
|
||||
// ###################################
|
||||
// ####################################
|
||||
// ## Section 2: GGML Op Definitions ##
|
||||
// ####################################
|
||||
|
||||
|
||||
// The following is an example showing the bare minimum for creating a test for a GGML op.
|
||||
|
|
@ -6222,7 +6222,7 @@ struct test_flash_attn_ext : public test_case {
|
|||
void initialize_tensors(ggml_context * ctx) override {
|
||||
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
|
||||
if (strcmp(t->name, "s") == 0) {
|
||||
// make the sink values more noticable in order to trigger a test failure when the implementation is wrong
|
||||
// make the sink values more noticeable in order to trigger a test failure when the implementation is wrong
|
||||
init_tensor_uniform(t, -10.0f, 10.0f);
|
||||
} else if (strcmp(t->name, "m") == 0) {
|
||||
init_tensor_kq_mask(t);
|
||||
|
|
|
|||
|
|
@ -438,7 +438,7 @@ static void test_templates(const struct common_chat_templates * tmpls, const std
|
|||
}
|
||||
|
||||
/**
|
||||
* Test if streaming=true is consistant with streaming=false for given partial parser
|
||||
* Test if streaming=true is consistent with streaming=false for given partial parser
|
||||
* Also test if there is any problem with partial message
|
||||
*/
|
||||
template <typename T>
|
||||
|
|
|
|||
|
|
@ -480,7 +480,7 @@ Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
|
|||
|
||||
Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
|
||||
|
||||
By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
|
||||
By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repetition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
|
||||
|
||||
Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
|
||||
|
||||
|
|
@ -531,7 +531,7 @@ These options help improve the performance and memory usage of the LLaMA models.
|
|||
|
||||
### NUMA support
|
||||
|
||||
- `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
|
||||
- `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
|
||||
- `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
|
||||
- `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitrary core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
|
||||
|
||||
|
|
|
|||
|
|
@ -110,7 +110,7 @@ struct callback_data {
|
|||
auto diff_filtered = filter_nonzero_rows(v_pos[il]);
|
||||
v_diff_filtered.push_back(diff_filtered);
|
||||
}
|
||||
return v_diff_filtered; // for convinient, we return the result std::vector
|
||||
return v_diff_filtered; // for convenient, we return the result std::vector
|
||||
}
|
||||
|
||||
// delete zero rows from a given 2D tensor
|
||||
|
|
|
|||
|
|
@ -95,4 +95,4 @@ Weighted averages of Σ(Act²), ZD Score and CosSim are also calculated.
|
|||
#### Important note on the computed Statistics
|
||||
|
||||
When using these statistics, please note that they are computed on the squared activations, **not on the actual (raw) activations**.
|
||||
Whilst the results are still useful, they're less realiable than using the raw values, and in the case of the cosine similarity, could be misleading if the tensor contains opposite vectors.
|
||||
Whilst the results are still useful, they're less reliable than using the raw values, and in the case of the cosine similarity, could be misleading if the tensor contains opposite vectors.
|
||||
|
|
|
|||
|
|
@ -68,7 +68,7 @@
|
|||
|
||||
#define TN_POS_EMBD "%s.position_embd.weight"
|
||||
#define TN_CLASS_EMBD "v.class_embd"
|
||||
#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backwrad compat
|
||||
#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backward compat
|
||||
#define TN_PATCH_EMBD_1 "v.patch_embd.weight.1"
|
||||
#define TN_PATCH_BIAS "v.patch_embd.bias"
|
||||
#define TN_NORM_EMBD "v.norm_embd.%s"
|
||||
|
|
|
|||
|
|
@ -46,7 +46,7 @@ struct clip_hparams {
|
|||
float image_std[3];
|
||||
|
||||
// for models using dynamic image size, we need to have a smaller image size to warmup
|
||||
// otherwise, user will get OOM everytime they load the model
|
||||
// otherwise, user will get OOM every time they load the model
|
||||
int32_t warmup_image_size = 0;
|
||||
int32_t warmup_audio_size = 3000;
|
||||
|
||||
|
|
@ -221,7 +221,7 @@ struct clip_model {
|
|||
// embeddings
|
||||
ggml_tensor * class_embedding = nullptr;
|
||||
ggml_tensor * patch_embeddings_0 = nullptr;
|
||||
ggml_tensor * patch_embeddings_1 = nullptr; // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL)
|
||||
ggml_tensor * patch_embeddings_1 = nullptr; // second Conv2D kernel when we decouple Conv3D along temporal dimension (Qwen2VL)
|
||||
ggml_tensor * patch_bias = nullptr;
|
||||
ggml_tensor * position_embeddings = nullptr;
|
||||
ggml_tensor * norm_embd_w = nullptr;
|
||||
|
|
|
|||
|
|
@ -2287,7 +2287,7 @@ static void normalize_image_u8_to_f32(const clip_image_u8 & src, clip_image_f32
|
|||
}
|
||||
}
|
||||
|
||||
// set of tools to manupulate images
|
||||
// set of tools to manipulate images
|
||||
// in the future, we can have HW acceleration by allowing this struct to access 3rd party lib like imagick or opencv
|
||||
struct img_tool {
|
||||
enum resize_algo {
|
||||
|
|
|
|||
|
|
@ -186,7 +186,7 @@ def trunc_normal_tf_(
|
|||
best when :math:`a \\leq \text{mean} \\leq b`.
|
||||
NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
|
||||
bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
|
||||
and the result is subsquently scaled and shifted by the mean and std args.
|
||||
and the result is subsequently scaled and shifted by the mean and std args.
|
||||
Args:
|
||||
tensor: an n-dimensional `torch.Tensor`
|
||||
mean: the mean of the normal distribution
|
||||
|
|
|
|||
|
|
@ -560,7 +560,7 @@ bool mtmd_audio_preprocessor_whisper::preprocess(const float * s
|
|||
for (size_t off = 0; off < (size_t) out_full.n_len; off += frames_per_chunk) {
|
||||
int n_len = std::min(frames_per_chunk, (size_t) out_full.n_len - off);
|
||||
if ((size_t) n_len < frames_per_chunk) {
|
||||
break; // last uncomplete chunk will always be a padded chunk, safe to ignore
|
||||
break; // last incomplete chunk will always be a padded chunk, safe to ignore
|
||||
}
|
||||
|
||||
mtmd_audio_mel out_chunk;
|
||||
|
|
|
|||
|
|
@ -27,10 +27,10 @@ In addition to the KL divergence the following statistics are calculated with `-
|
|||
* Ratio of mean FP16 PPL and quantized PPL. Uncertainty is estimated on logits, then propagated. The logarithm of this metric is also calculated and printed, it is 0 if the logit distributions are the same.
|
||||
* Difference of mean FP16 PPL and quantized PPL. Uncertainty is estimated on logits, then propagated.
|
||||
* Mean change in "correct" token probability. Positive values mean the model gets better at prediction, negative values mean it gets worse.
|
||||
* Pearson correlation coefficient of the "correct" token probabilites between models.
|
||||
* Pearson correlation coefficient of the "correct" token probabilities between models.
|
||||
* Percentiles of change in "correct" token probability. Positive values mean the model gets better at prediction, negative values mean it gets worse. Can be used to judge noise vs. quality loss from quantization. If the percentiles are symmetric then the quantization is essentially just adding noise. If the negative values are significantly larger than the positive values then this indicates that the model is actually becoming worse from the quantization.
|
||||
* The root mean square of the change in token probabilities. If you were to assume that the quantization simply causes Gaussian noise on the token probabilities then this would be the standard deviation of said noise. The uncertainty on the value is calculated that the change in token probabilities follows a Gaussian distribution. Related discussion: https://github.com/ggml-org/llama.cpp/discussions/2875 .
|
||||
* Same top p: Percentage of how often the token was assigned the highest probabilites by both models. The uncertainty is calculated from the Gaussian approximation of the binomial distribution.
|
||||
* Same top p: Percentage of how often the token was assigned the highest probabilities by both models. The uncertainty is calculated from the Gaussian approximation of the binomial distribution.
|
||||
|
||||
## LLaMA 3 8b Scoreboard
|
||||
|
||||
|
|
|
|||
|
|
@ -100,7 +100,7 @@ Examples:
|
|||
## Memory/Disk Requirements
|
||||
|
||||
When running the larger models, make sure you have enough disk space to store all the intermediate files.
|
||||
As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. For exmaple (Llama 3.1):
|
||||
As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. For example (Llama 3.1):
|
||||
|
||||
| Model | Original size | Quantized size (Q4_K_M) |
|
||||
| ----: | ------------: | ----------------------: |
|
||||
|
|
|
|||
|
|
@ -36,7 +36,7 @@
|
|||
|
||||
const params = signal({
|
||||
n_predict: 358, // 358 is a nice number
|
||||
temperature: 0.8, // adapt all following parameters to optimized min-p requierements. If for non-english, set to 0.6 or lower
|
||||
temperature: 0.8, // adapt all following parameters to optimized min-p requirements. If for non-english, set to 0.6 or lower
|
||||
repeat_last_n: 0, // 0 = disable penalty, -1 = context size
|
||||
repeat_penalty: 1.0, // 1.0 = disabled
|
||||
dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
|
||||
|
|
@ -108,7 +108,7 @@
|
|||
let importedTemplates = local_storage_getDataAsObject('user_templates')
|
||||
|
||||
if (importedTemplates) {
|
||||
// saved templates were successfuly imported.
|
||||
// saved templates were successfully imported.
|
||||
|
||||
console.log('Processing saved templates and updating default template')
|
||||
params.value = { ...params.value, image_data: [] };
|
||||
|
|
@ -129,7 +129,7 @@
|
|||
}
|
||||
|
||||
function userTemplateResetToDefault() {
|
||||
console.log('Reseting themplate to default')
|
||||
console.log('Reseting template to default')
|
||||
selectedUserTemplate.value.name = 'default';
|
||||
selectedUserTemplate.value.data = savedUserTemplates.value['default'];
|
||||
}
|
||||
|
|
|
|||
|
|
@ -63,7 +63,7 @@ export function trim_repeat_garbage_at_end(sIn, maxSubL=10, maxMatchLenThreshold
|
|||
|
||||
|
||||
/**
|
||||
* Simple minded logic to help remove repeating garbage at end of the string, till it cant.
|
||||
* Simple minded logic to help remove repeating garbage at end of the string, till it can't.
|
||||
* If its not able to trim, then it will try to skip a char at end and then trim, a few times.
|
||||
* This ensures that even if there are multiple runs of garbage with different patterns, the
|
||||
* logic still tries to munch through them.
|
||||
|
|
|
|||
|
|
@ -30,7 +30,7 @@ The UI follows a responsive web design so that the layout can adapt to available
|
|||
enough manner, in general.
|
||||
|
||||
Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool
|
||||
console. Parallely some of the directly useful to end-user settings can also be changed using the provided
|
||||
console. Parallelly some of the directly useful to end-user settings can also be changed using the provided
|
||||
settings ui.
|
||||
|
||||
NOTE: Current web service api doesnt expose the model context length directly, so client logic doesnt provide
|
||||
|
|
@ -38,7 +38,7 @@ any adaptive culling of old messages nor of replacing them with summary of their
|
|||
is a optional sliding window based chat logic, which provides a simple minded culling of old messages from
|
||||
the chat history before sending to the ai model.
|
||||
|
||||
NOTE: Wrt options sent with the request, it mainly sets temperature, max_tokens and optionaly stream for now.
|
||||
NOTE: Wrt options sent with the request, it mainly sets temperature, max_tokens and optionally stream for now.
|
||||
However if someone wants they can update the js file or equivalent member in gMe as needed.
|
||||
|
||||
NOTE: One may be able to use this to chat with openai api web-service /chat/completions endpoint, in a very
|
||||
|
|
@ -88,7 +88,7 @@ Once inside
|
|||
then the end user needs to enter the same.
|
||||
This keeps the logic simple, while still giving flexibility to the end user to
|
||||
manage any templating/tagging requirement wrt their messages to the model.
|
||||
* the logic doesnt insert newline at the begining and end wrt the prompt message generated.
|
||||
* the logic doesnt insert newline at the beginning and end wrt the prompt message generated.
|
||||
However if the chat being sent to /completions end point has more than one role's message,
|
||||
then insert newline when moving from one role's message to the next role's message, so
|
||||
that it can be clearly identified/distinguished.
|
||||
|
|
@ -101,8 +101,8 @@ Once inside
|
|||
Normally Completion mode doesnt need system prompt, while Chat mode can generate better/interesting
|
||||
responses with a suitable system prompt.
|
||||
* if chat.add_system_begin is used
|
||||
* you cant change the system prompt, after it is has been submitted once along with user query.
|
||||
* you cant set a system prompt, after you have submitted any user query
|
||||
* you can't change the system prompt, after it is has been submitted once along with user query.
|
||||
* you can't set a system prompt, after you have submitted any user query
|
||||
* if chat.add_system_anytime is used
|
||||
* one can change the system prompt any time during chat, by changing the contents of system prompt.
|
||||
* inturn the updated/changed system prompt will be inserted into the chat session.
|
||||
|
|
@ -129,7 +129,7 @@ Once inside
|
|||
|
||||
### Reason behind this
|
||||
|
||||
The idea is to be easy enough to use for basic purposes, while also being simple and easily discernable
|
||||
The idea is to be easy enough to use for basic purposes, while also being simple and easily discernible
|
||||
by developers who may not be from web frontend background (so inturn may not be familiar with template /
|
||||
end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things.
|
||||
|
||||
|
|
@ -167,7 +167,7 @@ It is attached to the document object. Some of these can also be updated using t
|
|||
messages that get inserted into prompt field wrt /Completion endpoint.
|
||||
|
||||
bTrimGarbage - whether garbage repeatation at the end of the generated ai response, should be
|
||||
trimmed or left as is. If enabled, it will be trimmed so that it wont be sent back as part of
|
||||
trimmed or left as is. If enabled, it will be trimmed so that it won't be sent back as part of
|
||||
subsequent chat history. At the same time the actual trimmed text is shown to the user, once
|
||||
when it was generated, so user can check if any useful info/data was there in the response.
|
||||
|
||||
|
|
@ -244,7 +244,7 @@ full chat history. This way if there is any response with garbage/repeatation, i
|
|||
mess with things beyond the next question/request/query, in some ways. The trim garbage
|
||||
option also tries to help avoid issues with garbage in the context to an extent.
|
||||
|
||||
Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space
|
||||
Set max_tokens to 1024, so that a relatively large previous response doesnt eat up the space
|
||||
available wrt next query-response. However dont forget that the server when started should
|
||||
also be started with a model context size of 1k or more, to be on safe side.
|
||||
|
||||
|
|
|
|||
|
|
@ -318,7 +318,7 @@ class SimpleChat {
|
|||
}
|
||||
|
||||
/**
|
||||
* Allow setting of system prompt, but only at begining.
|
||||
* Allow setting of system prompt, but only at beginning.
|
||||
* @param {string} sysPrompt
|
||||
* @param {string} msgTag
|
||||
*/
|
||||
|
|
@ -333,7 +333,7 @@ class SimpleChat {
|
|||
console.error(`ERRR:SimpleChat:SC:${msgTag}:You need to specify system prompt before any user query, ignoring...`);
|
||||
} else {
|
||||
if (this.xchat[0].content !== sysPrompt) {
|
||||
console.error(`ERRR:SimpleChat:SC:${msgTag}:You cant change system prompt, mid way through, ignoring...`);
|
||||
console.error(`ERRR:SimpleChat:SC:${msgTag}:You can't change system prompt, mid way through, ignoring...`);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -44,7 +44,7 @@ export function el_create_button(id, callback, name=undefined, innerText=undefin
|
|||
}
|
||||
|
||||
/**
|
||||
* Create a para and set it up. Optionaly append it to a passed parent.
|
||||
* Create a para and set it up. Optionally append it to a passed parent.
|
||||
* @param {string} text
|
||||
* @param {HTMLElement | undefined} elParent
|
||||
* @param {string | undefined} id
|
||||
|
|
@ -111,7 +111,7 @@ export function el_creatediv_boolbutton(id, label, texts, defaultValue, cb, clas
|
|||
/**
|
||||
* Create a select ui element, with a set of options to select from.
|
||||
* * options: an object which contains name-value pairs
|
||||
* * defaultOption: the value whose name should be choosen, by default.
|
||||
* * defaultOption: the value whose name should be chosen, by default.
|
||||
* * cb : the call back returns the name string of the option selected.
|
||||
*
|
||||
* @param {string} id
|
||||
|
|
|
|||
|
|
@ -57,7 +57,7 @@ To run a single test:
|
|||
./tests.sh unit/test_chat_completion.py::test_invalid_chat_completion_req
|
||||
```
|
||||
|
||||
Hint: You can compile and run test in single command, useful for local developement:
|
||||
Hint: You can compile and run test in single command, useful for local development:
|
||||
|
||||
```shell
|
||||
cmake --build build -j --target llama-server && ./tools/server/tests/tests.sh
|
||||
|
|
|
|||
Loading…
Reference in New Issue