// #define LOG_TARGET stderr
// #include "log.h"
//
-// The log target can also be redirected to a diffrent function
+// The log target can also be redirected to a different function
// like so:
//
-// #define LOG_TARGET log_handler_diffrent()
+// #define LOG_TARGET log_handler_different()
// #include "log.h"
//
-// FILE* log_handler_diffrent()
+// FILE* log_handler_different()
// {
// return stderr;
// }
// Disables logs entirely at runtime.
// Makes LOG() and LOG_TEE() produce no output,
-// untill enabled back.
+// until enabled back.
#define log_disable() log_disable_impl()
// INTERNAL, DO NOT USE
if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
# Transformers models put different tensors in different files, but
- # don't split indivdual tensors between files.
+ # don't split individual tensors between files.
model: LazyModel = {}
for mp in models_plus:
model.update(mp.model)
return func(*args)
CLASSES: dict[tuple[str, str], Any] = {
- # getattr used here as a workaround for mypy not being smart enough to detrmine
+ # getattr used here as a workaround for mypy not being smart enough to determine
# the staticmethods have a __func__ attribute.
('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
temp->ny = longer_side;
temp->size = 3 * longer_side * longer_side;
temp->data = new uint8_t[temp->size]();
- uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA
+ uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
// fill with background color
for (size_t i = 0; i < temp->size; i++) {
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
- This is a signficant percentage of your normal, say, 32K bpe vocab.
+ This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
# llama.cpp/examples/lookahead
-Demonstartion of lookahead decoding technique:
+Demonstration of lookahead decoding technique:
https://lmsys.org/blog/2023-11-21-lookahead-decoding/
}
if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
{
- return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimentional vector is not allowed", "size"), nullptr));
+ return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimensional vector is not allowed", "size"), nullptr));
}
std::vector<size_t> dim;
if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))
return content;
}
-// Call llama, return an event target that you can subcribe to
+// Call llama, return an event target that you can subscribe to
//
// Example:
//
cache_prompt: true
})
- /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
+ /* START: Support for storing prompt templates and parameters in browsers LocalStorage */
const local_storage_storageKey = "llamacpp_server_local_storage";
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: [] };
}
function userTemplateResetToDefault() {
- console.log('Reseting themplate to default')
+ console.log('Resetting template to default')
selectedUserTemplate.value.name = 'default';
selectedUserTemplate.value.data = savedUserTemplates.value['default'];
}
# llama.cpp/examples/speculative
-Demonstartion of speculative decoding and tree-based speculative decoding techniques
+Demonstration of speculative decoding and tree-based speculative decoding techniques
More info:
++n_past_tgt;
}
- // the first token is always proposed by the traget model before the speculation loop so we erase it here
+ // the first token is always proposed by the target model before the speculation loop so we erase it here
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
// ggml-backend v2 API
//
-// Seperate tensor and graph allocator objects
+// Separate tensor and graph allocator objects
// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
// The original API is kept as a wrapper around the new API
size_t vl = __riscv_vsetvl_e8m1(qk/2);
- // These tempory registers are for masking and shift operations
+ // These temporary registers are for masking and shift operations
vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
vl = 16;
- // retreive lane to multiply with scale
+ // retrieve lane to multiply with scale
vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
#define _USE_MATH_DEFINES // For M_PI on MSVC
#include "ggml-impl.h"
// we should just be careful :)
#pragma warning(disable: 4244 4267)
-// disable POSIX deprecation warnigns
+// disable POSIX deprecation warnings
// these functions are never going away, anyway
#pragma warning(disable: 4996)
#endif
static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
// WARN:
-// Mis-confguration can lead to problem that's hard to reason about:
+// Mis-configuration can lead to problem that's hard to reason about:
// * At best it crash or talks nosense.
// * At worst it talks slightly difference but hard to perceive.
//
GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
// view src0 and dst with these strides and data offset inbytes during acc
- // nb0 is implicitely element_size because src0 and dst are contiguous
+ // nb0 is implicitly element_size because src0 and dst are contiguous
size_t nb1 = ((int32_t *) dst->op_params)[0];
size_t nb2 = ((int32_t *) dst->op_params)[1];
size_t nb3 = ((int32_t *) dst->op_params)[2];
GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
// view src0 and dst with these strides and data offset inbytes during set
- // nb0 is implicitely element_size because src0 and dst are contiguous
+ // nb0 is implicitly element_size because src0 and dst are contiguous
size_t nb1 = ((int32_t *) dst->op_params)[0];
size_t nb2 = ((int32_t *) dst->op_params)[1];
size_t nb3 = ((int32_t *) dst->op_params)[2];
// insert new tensors recomputing src, reusing already made replacements,
// remember replacements: remember new tensors with mapping from corresponding gf nodes
// recurse for input tensors,
- // unless (i.e. terminating when) input tensors are replacments (like checkpoints)
+ // unless (i.e. terminating when) input tensors are replacements (like checkpoints)
node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
}
// insert rewritten backward node with replacements made into resulting backward graph gb
pip install build twine
```
-Then, folow these steps to release a new version:
+Then, follow these steps to release a new version:
1. Bump the version in `pyproject.toml`.
2. Build the package:
// The assumption is, since special tokens aren't meant to be exposed to end user, they are designed
// to be unmatchable by the tokenizer, therefore tokens from the vocab, which are unmatchable by the tokenizer
// are special tokens.
- // From testing, this appears to corelate 1:1 with special tokens.
+ // From testing, this appears to correlate 1:1 with special tokens.
//
// Counting special tokens and verifying in only one direction
const int64_t n_embd = hparams.n_embd;
const int64_t n_vocab = hparams.n_vocab;
- // helpers for smoother batch API transistion
+ // helpers for smoother batch API transition
// after deprecating the llama_eval calls, these will be removed
std::vector<llama_pos> pos;
// loop over the text
while (true) {
- // find the first occurence of a given special token in this fragment
+ // find the first occurrence of a given special token in this fragment
// passing offset argument only limit the "search area" but match coordinates
// are still relative to the source full raw_text
auto match = raw_text->find(special_token, raw_text_base_offset);
- // no occurences found, stop processing this fragment for a given special token
+ // no occurrences found, stop processing this fragment for a given special token
if (match == std::string::npos) break;
// check if match is within bounds of offset <-> length
}
// Min-heaps are used to efficiently collect the top-k elements (k=n_beams).
- // The repetative patterns below reflect the 2 stages of heaps:
+ // The repetitive patterns below reflect the 2 stages of heaps:
// * Gather elements until the vector is full, then call std::make_heap() on it.
// * If the heap is full and a new element is found that should be included, pop the
// least element to the back(), replace it with the new, then push it into the heap.
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
#include "ggml.h"
#include <cmath>
printf(" --size SIZE set test size, divisible by 32 (L1_SIZE:%d)\n", L1_SIZE);
printf(" -3 use size as L1, L2, L3 sizes (L1:%d L2:%d L3:%d)\n", L1_SIZE, L2_SIZE, L3_SIZE);
printf(" -4 use size as L1, L2, L3, MEM sizes (L1:%d L2:%d L3:%d MEM:%d)\n", L1_SIZE, L2_SIZE, L3_SIZE, MEM_SIZE);
- printf(" --op OP set test opration as quantize_row_q_reference, quantize_row_q, dequantize_row_q,\n");
+ printf(" --op OP set test operation as quantize_row_q_reference, quantize_row_q, dequantize_row_q,\n");
printf(" quantize_row_q_dot, vec_dot_q (all)\n");
printf(" --type TYPE set test type as");
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
}
int alignment = std::stoi(argv[i]);
if (alignment < 0 || alignment > MAX_ALIGNMENT) {
- fprintf(stderr, "error: aligment-offset must be less than %d\n", MAX_ALIGNMENT);
+ fprintf(stderr, "error: alignment-offset must be less than %d\n", MAX_ALIGNMENT);
invalid_param = true;
break;
}
overview / pkg / ggml / sources / llama.cpp / commitdiff