llama.cpp/src/llama-quant.cpp

1863 lines
82 KiB
C++
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#include "llama-quant.h"
#include "llama-impl.h"
#include "llama-model.h"
#include "llama-model-loader.h"
#include <algorithm>
#include <atomic>
#include <cmath>
#include <cstring>
#include <cinttypes>
#include <fstream>
#include <mutex>
#include <random>
#include <regex>
#include <thread>
#include <unordered_map>
// Quantization types. Changes to this struct must be replicated in quantize.cpp
struct tensor_quantization {
std::string name;
ggml_type quant = GGML_TYPE_COUNT;
};
static bool is_iq(const enum ggml_type t) {
switch (t) {
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
return true;
default:
return false;
}
}
static bool is_iq(const enum llama_ftype t) {
switch (t) {
case LLAMA_FTYPE_MOSTLY_IQ1_S:
case LLAMA_FTYPE_MOSTLY_IQ1_M:
case LLAMA_FTYPE_MOSTLY_IQ2_XXS:
case LLAMA_FTYPE_MOSTLY_IQ2_XS:
case LLAMA_FTYPE_MOSTLY_IQ2_S:
case LLAMA_FTYPE_MOSTLY_IQ2_M:
case LLAMA_FTYPE_MOSTLY_IQ3_XXS:
case LLAMA_FTYPE_MOSTLY_IQ3_XS:
case LLAMA_FTYPE_MOSTLY_IQ3_S:
case LLAMA_FTYPE_MOSTLY_IQ3_M:
case LLAMA_FTYPE_MOSTLY_IQ4_XS:
case LLAMA_FTYPE_MOSTLY_IQ4_NL:
return true;
default:
return false;
}
}
static enum ggml_type fallback_type(const enum ggml_type new_type) {
switch (new_type) {
case GGML_TYPE_TQ1_0:
case GGML_TYPE_TQ2_0:
return GGML_TYPE_Q4_0; // symmetric-ish fallback
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_IQ4_XS:
return GGML_TYPE_IQ4_NL;
case GGML_TYPE_Q4_K:
return GGML_TYPE_Q5_0;
case GGML_TYPE_Q5_K:
return GGML_TYPE_Q5_1;
case GGML_TYPE_Q6_K:
return GGML_TYPE_Q8_0;
default:
return new_type;
}
}
static void zeros(std::ofstream & file, size_t n) {
char zero = 0;
for (size_t i = 0; i < n; ++i) {
file.write(&zero, 1);
}
}
static std::string remap_layer(const std::string & orig_name, const std::vector<int> & prune, std::map<int, std::string> & mapped, int & next_id) {
if (prune.empty()) {
return orig_name;
}
static const std::regex pattern(R"(blk\.(\d+)\.)");
if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
const int blk = std::stoi(match[1]);
std::string new_name = orig_name;
if (mapped.count(blk)) {
// Already mapped, do nothing
} else if (std::find(prune.begin(), prune.end(), blk) != prune.end()) {
mapped[blk] = "";
} else if (blk < prune.front()) {
mapped[blk] = std::to_string(blk);
next_id = blk + 1;
} else {
mapped[blk] = std::to_string(next_id);
++next_id;
}
return mapped[blk].empty() ? mapped[blk] : new_name.replace(match.position(1), match.length(1), mapped[blk]);
}
return orig_name;
}
static std::string remap_imatrix (const std::string & orig_name, const std::map<int, std::string> & mapped) {
if (mapped.empty()) {
return orig_name;
}
static const std::regex pattern(R"(blk\.(\d+)\.)");
if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
const std::string blk(match[1]);
std::string new_name = orig_name;
for (const auto & p : mapped) {
if (p.second == blk) {
return new_name.replace(match.position(1), match.length(1), std::to_string(p.first));
}
}
GGML_ABORT("\n%s: imatrix mapping error for %s\n", __func__, orig_name.c_str());
}
return orig_name;
}
struct quantize_state_impl {
const llama_model & model;
const llama_model_quantize_params * params;
int n_attention_wv = 0;
int n_ffn_down = 0;
int n_ffn_gate = 0;
int n_ffn_up = 0;
int i_attention_wv = 0;
int i_ffn_down = 0;
int i_ffn_gate = 0;
int i_ffn_up = 0;
int n_k_quantized = 0;
int n_fallback = 0;
bool has_imatrix = false;
bool has_activations = false;
// used to figure out if a model shares tok_embd with the output weight
bool has_output = false;
quantize_state_impl(const llama_model & model, const llama_model_quantize_params * params)
: model(model)
, params(params)
{}
};
static void llama_tensor_dequantize_impl(
ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
const size_t nelements, const int nthread
) {
if (output.size() < nelements) {
output.resize(nelements);
}
float * f32_output = (float *) output.data();
const ggml_type_traits * qtype = ggml_get_type_traits(tensor->type);
if (ggml_is_quantized(tensor->type)) {
if (qtype->to_float == NULL) {
throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
}
} else if (tensor->type != GGML_TYPE_F16 &&
tensor->type != GGML_TYPE_BF16) {
throw std::runtime_error(format("cannot dequantize/convert tensor type %s", ggml_type_name(tensor->type)));
}
if (nthread < 2) {
if (tensor->type == GGML_TYPE_F16) {
ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor->data, f32_output, nelements);
} else if (tensor->type == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
} else if (ggml_is_quantized(tensor->type)) {
qtype->to_float(tensor->data, f32_output, nelements);
} else {
GGML_ABORT("fatal error"); // unreachable
}
return;
}
size_t block_size;
if (tensor->type == GGML_TYPE_F16 ||
tensor->type == GGML_TYPE_BF16) {
block_size = 1;
} else {
block_size = (size_t)ggml_blck_size(tensor->type);
}
size_t block_size_bytes = ggml_type_size(tensor->type);
GGML_ASSERT(nelements % block_size == 0);
size_t nblocks = nelements / block_size;
size_t blocks_per_thread = nblocks / nthread;
size_t spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count
size_t in_buff_offs = 0;
size_t out_buff_offs = 0;
for (int tnum = 0; tnum < nthread; tnum++) {
size_t thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread
size_t thr_elems = thr_blocks * block_size; // number of elements for this thread
size_t thr_block_bytes = thr_blocks * block_size_bytes; // number of input bytes for this thread
auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) {
if (typ == GGML_TYPE_F16) {
ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
} else if (typ == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
} else {
qtype->to_float(inbuf, outbuf, nels);
}
};
workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
in_buff_offs += thr_block_bytes;
out_buff_offs += thr_elems;
}
for (auto & w : workers) { w.join(); }
workers.clear();
}
static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
const std::string name = ggml_get_name(tensor);
// TODO: avoid hardcoded tensor names - use the TN_* constants
const llm_arch arch = qs.model.arch;
const auto tn = LLM_TN(arch);
auto use_more_bits = [](int i_layer, int n_layers) -> bool {
return i_layer < n_layers/8 || i_layer >= 7*n_layers/8 || (i_layer - n_layers/8)%3 == 2;
};
const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) {
if (n_expert > 1) {
// Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but occasionally randomly
// sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
// for getting the current layer as I initially thought, and we need to resort to parsing the
// tensor name.
if (sscanf(name, "blk.%d.", &i_layer) != 1) {
throw std::runtime_error(format("Failed to determine layer for tensor %s", name));
}
if (i_layer < 0 || i_layer >= n_layer) {
throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer));
}
}
return std::make_pair(i_layer, n_layer);
};
// for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
// with the quantization of the output tensor
if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
new_type = qs.params->output_tensor_type;
} else {
const int64_t nx = tensor->ne[0];
const int64_t qk_k = ggml_blck_size(new_type);
if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
new_type = GGML_TYPE_Q8_0;
}
else if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
new_type = GGML_TYPE_Q5_K;
}
else if (new_type != GGML_TYPE_Q8_0) {
new_type = GGML_TYPE_Q6_K;
}
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
// MoE tensors -> MXFP4
// other tensors -> Q8_0
if (tensor->ne[2] > 1) {
new_type = GGML_TYPE_MXFP4;
} else {
new_type = GGML_TYPE_Q8_0;
}
} else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") {
if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
new_type = qs.params->token_embedding_type;
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
new_type = GGML_TYPE_Q2_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
new_type = GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
new_type = GGML_TYPE_Q4_K;
}
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
if (name.find("attn_v.weight") != std::string::npos) {
if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
else new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
++qs.i_attention_wv;
}
else if (qs.model.hparams.n_expert == 8 && name.find("attn_k.weight") != std::string::npos) {
new_type = GGML_TYPE_Q4_K;
}
else if (name.find("ffn_down") != std::string::npos) {
if (qs.i_ffn_down < qs.n_ffn_down/8) {
new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
}
++qs.i_ffn_down;
}
else if (name.find("attn_output.weight") != std::string::npos) {
if (qs.model.hparams.n_expert == 8) {
new_type = GGML_TYPE_Q5_K;
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_S;
}
}
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q5_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
if (qs.model.type == LLM_TYPE_70B) {
// In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
// nearly negligible increase in model size by quantizing this tensor with more bits:
if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
}
if (qs.model.hparams.n_expert == 8) {
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
++qs.i_attention_wv;
} else if (name.find("attn_k.weight") != std::string::npos) {
if (qs.model.hparams.n_expert == 8) {
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ2_S;
}
} else if (name.find("attn_q.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ2_S;
}
} else if (name.find("ffn_down") != std::string::npos) {
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) {
new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
: arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
: GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (i_layer < n_layer/8 ||
(qs.model.hparams.n_expert == 8 && use_more_bits(i_layer, n_layer)))) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
if (arch == LLM_ARCH_FALCON) {
new_type = i_layer < n_layer/16 ? GGML_TYPE_Q6_K :
use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
} else {
if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
}
}
else if (i_layer < n_layer/8 && (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && !qs.has_imatrix) {
new_type = GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
new_type = GGML_TYPE_Q5_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0)
&& qs.has_imatrix && i_layer < n_layer/8) {
// Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
// We only do it when an imatrix is provided because a) we want to make sure that one can always get the
// same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
}
++qs.i_ffn_down;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (arch != LLM_ARCH_FALCON) {
if (qs.model.hparams.n_expert == 8) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL ||
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) {
new_type = GGML_TYPE_Q5_K;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M ) new_type = GGML_TYPE_Q4_K;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
}
}
else if (name.find("attn_qkv.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate") != std::string::npos) {
auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
new_type = GGML_TYPE_IQ3_XXS;
}
++qs.i_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
new_type = GGML_TYPE_IQ3_XXS;
}
++qs.i_ffn_up;
}
// if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
//}
// IK: let's remove this, else Q2_K is almost the same as Q3_K_S
//else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
//}
// This can be used to reduce the size of the Q5_K_S model.
// The associated PPL increase is fully in line with the size reduction
//else {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
//}
bool convert_incompatible_tensor = false;
{
const int64_t nx = tensor->ne[0];
const int64_t ny = tensor->ne[1];
const int64_t qk_k = ggml_blck_size(new_type);
if (nx % qk_k != 0) {
LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
convert_incompatible_tensor = true;
} else {
++qs.n_k_quantized;
}
}
if (convert_incompatible_tensor) {
switch (new_type) {
case GGML_TYPE_TQ1_0:
case GGML_TYPE_TQ2_0: new_type = GGML_TYPE_Q4_0; break; // TODO: use a symmetric type instead
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q5_1; break;
case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break;
default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
}
if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
new_type = GGML_TYPE_F16;
}
LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
++qs.n_fallback;
}
return new_type;
}
static size_t llama_tensor_quantize_impl(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
if (nthread < 2) {
// single-thread
size_t new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix);
if (!ggml_validate_row_data(new_type, new_data, new_size)) {
throw std::runtime_error("quantized data validation failed");
}
return new_size;
}
std::mutex mutex;
int64_t counter = 0;
size_t new_size = 0;
bool valid = true;
auto compute = [&mutex, &counter, &new_size, &valid, new_type, f32_data, new_data, chunk_size,
nrows, n_per_row, imatrix]() {
const int64_t nrows_per_chunk = chunk_size / n_per_row;
size_t local_size = 0;
while (true) {
std::unique_lock<std::mutex> lock(mutex);
int64_t first_row = counter; counter += nrows_per_chunk;
if (first_row >= nrows) {
if (local_size > 0) {
new_size += local_size;
}
break;
}
lock.unlock();
const int64_t this_nrow = std::min(nrows - first_row, nrows_per_chunk);
size_t this_size = ggml_quantize_chunk(new_type, f32_data, new_data, first_row * n_per_row, this_nrow, n_per_row, imatrix);
local_size += this_size;
// validate the quantized data
const size_t row_size = ggml_row_size(new_type, n_per_row);
void * this_data = (char *) new_data + first_row * row_size;
if (!ggml_validate_row_data(new_type, this_data, this_size)) {
std::unique_lock<std::mutex> lock(mutex);
valid = false;
break;
}
}
};
for (int it = 0; it < nthread - 1; ++it) {
workers.emplace_back(compute);
}
compute();
for (auto & w : workers) { w.join(); }
workers.clear();
if (!valid) {
throw std::runtime_error("quantized data validation failed");
}
return new_size;
}
// Returns per-tensor type overrides to meet target BPW at lowest ppl
static std::unordered_map<std::string, ggml_type> target_bpw_type(
llama_model_loader & ml,
std::vector<no_init<uint8_t>> & buffer,
const llama_model & model,
const std::vector<const llama_model_loader::llama_tensor_weight *> & tensors,
const std::map<int, std::string> & mapped,
const std::unordered_map<std::string, std::vector<float>> * values_data,
const std::unordered_map<std::string, std::vector<float>> * activations_data,
const llama_model_quantize_params * params,
int nthread
) {
struct candidate_types {
ggml_type type;
float bpw;
size_t bytes;
float error;
};
struct tensor_info {
const llama_model_loader::llama_tensor_weight * w = nullptr;
std::vector<candidate_types> candidate = {};
int choice = -1;
float min_bpw = 0.0;
float max_bpw = 0.0;
size_t n_elements = 0;
};
constexpr ggml_type k_quants[] = {
GGML_TYPE_Q2_K,
GGML_TYPE_Q3_K,
GGML_TYPE_Q4_K,
GGML_TYPE_Q5_K,
GGML_TYPE_Q6_K,
GGML_TYPE_Q8_0,
// TODO: find better way to handle F16/BF16
#ifdef GGML_USE_METAL
GGML_TYPE_F16
#else
GGML_TYPE_BF16
#endif
};
constexpr ggml_type iq_quants[] = {
GGML_TYPE_IQ1_S,
GGML_TYPE_IQ2_S,
GGML_TYPE_IQ3_S,
GGML_TYPE_IQ4_XS,
GGML_TYPE_Q5_K,
GGML_TYPE_Q6_K,
GGML_TYPE_Q8_0
};
auto tensor_bytes = [](const ggml_tensor * t, const ggml_type typ) -> size_t {
const int64_t n_per_row = t->ne[0];
const size_t row_sz = ggml_row_size(typ, n_per_row);
const int64_t nrows = ggml_nrows(t);
return (size_t)nrows * row_sz;
};
auto tensor_bpw = [&](const ggml_tensor * t, const ggml_type typ) -> double {
const int64_t nelem = ggml_nelements(t);
const size_t bytes = tensor_bytes(t, typ);
return (double)bytes * 8.0 / (double)nelem;
};
auto is_compatible = [&](const ggml_tensor * t, const ggml_type typ) -> bool {
const int64_t n_per_row = t->ne[0];
const int64_t blck = ggml_blck_size(typ);
if (blck <= 1) { return true; }
return n_per_row % blck == 0;
};
auto make_compatible = [&](const ggml_tensor * t, const ggml_type typ) -> ggml_type {
if (is_compatible(t, typ)) { return typ; }
ggml_type fb = fallback_type(typ);
if (is_compatible(t, fb)) { return fb; }
return GGML_TYPE_F16;
};
auto name_tn = LLM_TN(model.arch);
auto can_quantize = [&](const ggml_tensor * t) -> bool {
// This list should be kept in sync with llama_tensor_quantize_impl()
const std::string name = ggml_get_name(t);
bool q = name.rfind("weight") == name.size() - 6;
q &= ggml_n_dims(t) >= 2;
q &= name.find("_norm.weight") == std::string::npos;
q &= name.find("ffn_gate_inp.weight") == std::string::npos;
q &= name.find("altup") == std::string::npos;
q &= name.find("laurel") == std::string::npos;
q &= name.find("per_layer_model_proj") == std::string::npos;
q &= name != name_tn(LLM_TENSOR_POS_EMBD, "weight");
q &= name != name_tn(LLM_TENSOR_TOKEN_TYPES, "weight");
q &= name.find("ssm_conv1d.weight") == std::string::npos;
q &= name.find("shortconv.conv.weight") == std::string::npos;
q &= name.find("time_mix_first.weight") == std::string::npos;
q &= name.find("time_mix_w0.weight") == std::string::npos;
q &= name.find("time_mix_w1.weight") == std::string::npos;
q &= name.find("time_mix_w2.weight") == std::string::npos;
q &= name.find("time_mix_v0.weight") == std::string::npos;
q &= name.find("time_mix_v1.weight") == std::string::npos;
q &= name.find("time_mix_v2.weight") == std::string::npos;
q &= name.find("time_mix_a0.weight") == std::string::npos;
q &= name.find("time_mix_a1.weight") == std::string::npos;
q &= name.find("time_mix_a2.weight") == std::string::npos;
q &= name.find("time_mix_g1.weight") == std::string::npos;
q &= name.find("time_mix_g2.weight") == std::string::npos;
q &= name.find("time_mix_decay_w1.weight") == std::string::npos;
q &= name.find("time_mix_decay_w2.weight") == std::string::npos;
q &= name.find("time_mix_lerp_fused.weight") == std::string::npos;
q &= name.find("attn_rel_b.weight") == std::string::npos;
q &= !params->only_copy;
return q;
};
// Estimate error for a given type using a sampled subset of rows
auto estimate_error = [&](const ggml_tensor * t,
const ggml_type quant_type,
const std::vector<float> & f32_sample,
const std::vector<int64_t> & sample_rows_per_slice,
const float * values_sample,
const float * activations_sample,
std::vector<uint8_t> & quantized_buffer,
std::vector<float> & dequantized_buffer) -> double
{
const int64_t n_per_row = t->ne[0];
const int64_t nrows = t->ne[1];
const int64_t ne2 = t->ne[2] > 0 ? t->ne[2] : 1;
const size_t sample_element_count = f32_sample.size();
const size_t sample_row_count = sample_element_count / (size_t)n_per_row;
if (sample_row_count == 0) { return 0.0; }
const size_t row_sz = ggml_row_size(quant_type, n_per_row);
const size_t buffer_sz = row_sz * sample_row_count;
if (quantized_buffer.size() < buffer_sz) { quantized_buffer.resize(buffer_sz); }
if (dequantized_buffer.size() < sample_element_count) { dequantized_buffer.resize(sample_element_count); }
const bool has_values = values_sample != nullptr;
const bool has_activations = activations_sample != nullptr;
// Bias denominators per slice (only needed if we have activations)
std::vector<double> bias_denominator_per_slice(ne2, 0.0);
if (has_activations) {
for (int64_t s = 0; s < ne2; ++s) {
const float * values = has_values ? values_sample + s * n_per_row : nullptr;
const float * activations = activations_sample + s * n_per_row;
double denom = 0.0;
for (int64_t j = 0; j < n_per_row; ++j) {
const double a = activations[j];
const double w = values ? values[j] : 1.0;
denom += w * a * a;
}
bias_denominator_per_slice[s] = denom;
}
}
// Compute per-row squared norms with weighting (if values are provided)
std::vector<double> row_sq_norm(sample_row_count, 0.0);
{
size_t offset = 0;
size_t row_idx = 0;
for (int64_t s = 0; s < ne2; ++s) {
const int64_t rs = sample_rows_per_slice[s];
if (rs == 0) { continue; }
const float * values = has_values ? values_sample + s * n_per_row : nullptr;
for (int64_t r = 0; r < rs; ++r, ++row_idx) {
const float * x = f32_sample.data() + offset;
double rsn = 0.0;
if (values) {
for (int64_t j = 0; j < n_per_row; ++j) {
const double v = values[j];
const double xx = x[j];
rsn += v * xx * xx;
}
} else {
for (int64_t j = 0; j < n_per_row; ++j) {
const double xx = x[j];
rsn += xx * xx;
}
}
row_sq_norm[row_idx] = rsn;
offset += (size_t)n_per_row;
}
}
}
// Quantize sampled rows slice-by-slice into quantized_buffer
{
size_t q_offset = 0;
size_t f_offset = 0;
for (int64_t slice = 0; slice < ne2; ++slice) {
const int64_t rs = sample_rows_per_slice[slice];
if (rs == 0) { continue; }
const float * value = has_values ? values_sample + slice * n_per_row : nullptr;
(void)ggml_quantize_chunk(quant_type, f32_sample.data() + f_offset, quantized_buffer.data() + q_offset, 0, rs, n_per_row, value);
q_offset += row_sz * (size_t)rs;
f_offset += (size_t)rs * (size_t)n_per_row;
}
}
// Dequantize into dequantized_buffer
{
const ggml_type_traits * traits = ggml_get_type_traits(quant_type);
auto row_to_float = [&](size_t r) {
uint8_t * src = quantized_buffer.data() + r * row_sz;
float * dst = dequantized_buffer.data() + r * (size_t)n_per_row;
if (quant_type == GGML_TYPE_F16) {
ggml_fp16_to_fp32_row((const ggml_fp16_t *)src, dst, (int)n_per_row);
} else if (quant_type == GGML_TYPE_BF16) {
ggml_bf16_to_fp32_row((const ggml_bf16_t *)src, dst, (int)n_per_row);
} else {
if (!traits || !traits->to_float) {
LLAMA_LOG_WARN("%s: unsupported quantization type %s\n", __func__, ggml_type_name(quant_type));
return false;
}
traits->to_float(src, dst, (int)n_per_row);
}
return true;
};
for (size_t r = 0; r < sample_row_count; ++r) {
if (!row_to_float(r)) { return 1e35; }
}
}
// Compute error
size_t offset = 0;
size_t row_idx = 0;
double total_err = 0.0;
for (int64_t slice = 0; slice < ne2; ++slice) {
const int64_t rs = sample_rows_per_slice[slice];
if (rs == 0) { continue; }
const float * values = has_values ? values_sample + slice * n_per_row : nullptr;
const float * activations = has_activations ? activations_sample + slice * n_per_row : nullptr;
const double bias_denom = has_activations ? bias_denominator_per_slice[slice] : 0.0;
double slice_err = 0.0;
for (int64_t r = 0; r < rs; ++r, ++row_idx) {
const float * x = f32_sample.data() + offset;
const float * y = dequantized_buffer.data() + offset;
double weighted_mse = 0.0;
double bias_num = 0.0;
if (values && activations) {
for (int64_t j = 0; j < n_per_row; ++j) {
const double v = values[j];
const double e = y[j] - x[j];
const double a = activations[j];
weighted_mse += v * e * e;
bias_num += v * e * a;
}
} else if (values) {
for (int64_t j = 0; j < n_per_row; ++j) {
const double v = values[j];
const double e = y[j] - x[j];
weighted_mse += v * e * e;
}
} else if (activations) {
for (int64_t j = 0; j < n_per_row; ++j) {
const double e = y[j] - x[j];
const double a = activations[j];
weighted_mse += e * e;
bias_num += e * a;
}
} else {
for (int64_t j = 0; j < n_per_row; ++j) {
const double e = y[j] - x[j];
weighted_mse += e * e;
}
}
// bias_lambda adjusts the trade-off between systematic bias (introduced by blockwise scaling) and MSE
// larger value favours quantisation types that produce smaller bias even if the MSE is slightly larger
constexpr float bias_lambda = 1.5f;
constexpr double epsilon = 1e-12;
double err_num = weighted_mse;
if (activations && bias_lambda != 0.0f) {
const double proj = bias_num * bias_num / (bias_denom + epsilon);
err_num += (double)bias_lambda * proj;
}
const double err_den = row_sq_norm[row_idx] + epsilon;
slice_err += err_num / err_den;
offset += (size_t)n_per_row;
}
const double scale_rows = (double)nrows / std::max(1.0, (double)rs);
total_err += slice_err * scale_rows;
}
return std::isfinite(total_err) ? total_err : 1e35;
};
std::vector<tensor_info> all;
all.reserve(tensors.size());
for (const auto * tw : tensors) {
std::vector<std::thread> workers;
workers.reserve(std::max(1, nthread));
ggml_tensor * t = tw->tensor;
const std::string name = ggml_get_name(t);
if (!can_quantize(t)) { continue; }
LLAMA_LOG_INFO("\t%s: - processing tensor %45s \t(%12d elements)\n", __func__, name.c_str(), (int)ggml_nelements(t));
if (!ml.use_mmap) {
if (buffer.size() < ggml_nbytes(t)) { buffer.resize(ggml_nbytes(t)); }
t->data = buffer.data();
}
ml.load_data_for(t);
// Dequantize only sampled rows into f32_sample
const int64_t n_per_row = t->ne[0];
const int64_t nrows_total = t->ne[1];
const int64_t ne2 = t->ne[2] > 0 ? t->ne[2] : 1;
// Larger sample_rows_per_expert values may result in more accurate error estimates, but will take longer to compute
constexpr int sample_rows_per_expert = 384;
std::vector<float> f32_sample;
f32_sample.reserve((size_t)ne2 * (size_t)std::min<int64_t>(nrows_total, sample_rows_per_expert) * (size_t)n_per_row);
// deterministic sampling seed based on tensor name + fixed constant
std::mt19937 rng(std::hash<std::string>{}(name) ^0xeabada55cafed00d);
std::vector<int64_t> sample_rows_per_slice(ne2, 0);
const int64_t sample_rows_max = std::max<int64_t>(1, std::min<int64_t>(nrows_total, sample_rows_per_expert));
const int64_t stride = std::max<int64_t>(1, nrows_total / sample_rows_max);
std::vector<float> row_buffer(n_per_row);
const ggml_type src_type = t->type;
const ggml_type_traits *src_traits = ggml_get_type_traits(src_type);
const bool src_is_quant = ggml_is_quantized(src_type);
const size_t src_row_sz = ggml_row_size(src_type, n_per_row);
for (int64_t slice = 0; slice < ne2; ++slice) {
int64_t current_sampled_rows = 0;
int64_t offset = 0;
if (stride > 1) {
std::uniform_int_distribution<int64_t> dist(0, stride - 1);
offset = dist(rng);
}
for (int64_t r = offset; r < nrows_total && current_sampled_rows < sample_rows_max; r += stride) {
if (src_type == GGML_TYPE_F32) {
const float * src_row = (const float *)t->data + slice * (n_per_row * nrows_total) + r * n_per_row;
f32_sample.insert(f32_sample.end(), src_row, src_row + n_per_row);
} else if (src_type == GGML_TYPE_F16) {
const ggml_fp16_t * src_row = (const ggml_fp16_t *)((const uint8_t *)t->data + slice * (src_row_sz * nrows_total) + r * src_row_sz);
ggml_fp16_to_fp32_row(src_row, row_buffer.data(), (int)n_per_row);
f32_sample.insert(f32_sample.end(), row_buffer.begin(), row_buffer.end());
} else if (src_type == GGML_TYPE_BF16) {
const ggml_bf16_t * src_row = (const ggml_bf16_t *)((const uint8_t *)t->data + slice * (src_row_sz * nrows_total) + r * src_row_sz);
ggml_bf16_to_fp32_row(src_row, row_buffer.data(), (int)n_per_row);
f32_sample.insert(f32_sample.end(), row_buffer.begin(), row_buffer.end());
} else if (src_is_quant) {
const uint8_t * qrow = (const uint8_t *)t->data + slice * (src_row_sz * nrows_total) + r * src_row_sz;
if (!src_traits || !src_traits->to_float) {
throw std::runtime_error(format("cannot dequantize type %s for sampling", ggml_type_name(src_type)));
}
src_traits->to_float(qrow, row_buffer.data(), (int)n_per_row);
f32_sample.insert(f32_sample.end(), row_buffer.begin(), row_buffer.end());
} else {
throw std::runtime_error(format("unsupported src type %s for sampling", ggml_type_name(src_type)));
}
++current_sampled_rows;
}
sample_rows_per_slice[slice] = current_sampled_rows;
}
auto side_data = [&](const std::unordered_map<std::string, std::vector<float>> * m, const std::string & tensor_name) -> std::pair<const float*, size_t> {
if (!m) { return {nullptr, 0}; }
const std::string key = remap_imatrix(tensor_name, mapped);
const auto it = m->find(key);
if (it == m->end()) { return {nullptr, 0}; }
return { it->second.data(), it->second.size() };
};
// Copy this row's side data (values and activations), or broadcasts to all slices
auto copy_or_broadcast = [&](const float *src, size_t src_sz, std::vector<float> &dst) {
const size_t want = (size_t)ne2 * (size_t)n_per_row;
dst.clear();
if (!src || src_sz == 0) { return; }
if (src_sz == want) {
dst.resize(want);
std::memcpy(dst.data(), src, want * sizeof(float));
} else if (src_sz == (size_t)n_per_row) {
dst.resize(want);
for (int64_t s = 0; s < ne2; ++s) {
std::memcpy(dst.data() + s * n_per_row, src, n_per_row * sizeof(float));
}
} else {
LLAMA_LOG_WARN("%s: side data size mismatch for %s: got %zu, expected %zu or %zu; ignoring\n",
__func__, name.c_str(), src_sz, (size_t)n_per_row, want);
}
};
const auto [values_all, values_sz] = side_data(values_data, name);
const auto [activations_all, activations_sz] = side_data(activations_data, name);
std::vector<float> values_sample;
std::vector<float> activations_sample;
if (values_all) { copy_or_broadcast(values_all, values_sz, values_sample); }
if (activations_all) { copy_or_broadcast(activations_all, activations_sz, activations_sample); }
const int64_t nelem = ggml_nelements(t);
tensor_info info;
info.w = tw;
info.n_elements = nelem;
// Prepare scratch buffers sized for the largest candidate row size
size_t total_sampled_rows = f32_sample.size() / n_per_row;
// Build list of candidate types first (compatible ones)
const ggml_type * base_arr = is_iq(params->ftype) ? iq_quants : k_quants;
const size_t base_sz = is_iq(params->ftype) ? std::size(iq_quants) : std::size(k_quants);
size_t max_row_sz = 0;
const bool has_valid_imatrix = !values_sample.empty() && values_sample.size() == (size_t)ne2 * (size_t)n_per_row;
std::vector<ggml_type> compatible_candidates;
compatible_candidates.reserve(base_sz);
for (size_t i = 0; i < base_sz; ++i) {
ggml_type ts_type = base_arr[i];
if (is_iq(ts_type) && !has_valid_imatrix) {
LLAMA_LOG_WARN("%s: skipping %s quantization for %s, no or mismatched imatrix provided\n",
__func__, ggml_type_name(ts_type), name.c_str());
continue;
}
ggml_type tt = make_compatible(t, ts_type);
if (!is_compatible(t, tt)) { continue; }
compatible_candidates.push_back(tt);
max_row_sz = std::max(max_row_sz, ggml_row_size(tt, n_per_row));
}
std::sort(compatible_candidates.begin(), compatible_candidates.end());
compatible_candidates.erase(std::unique(compatible_candidates.begin(), compatible_candidates.end()), compatible_candidates.end());
// Now evaluate candidates
std::vector<candidate_types> eval_candidates(compatible_candidates.size());
const float * values = values_sample.empty() ? nullptr : values_sample.data();
const float * activations = activations_sample.empty() ? nullptr : activations_sample.data();
std::vector<uint8_t> quantized_buffer(max_row_sz * total_sampled_rows);
std::vector<float> dequantised_buffer(f32_sample.size());
int n_eval_threads = std::max(1, std::min<int>(nthread, (int)compatible_candidates.size()));
std::atomic<size_t> cidx{0};
std::vector<std::thread> eval_workers;
eval_workers.reserve(n_eval_threads);
for (int ti = 0; ti < n_eval_threads; ++ti) {
eval_workers.emplace_back([&] {
// thread-local scratch
std::vector<uint8_t> tl_quantized_buffer(quantized_buffer.size());
std::vector<float> tl_dequantised_buffer(dequantised_buffer.size());
for (;;) {
const size_t i = cidx.fetch_add(1, std::memory_order_relaxed);
if (i >= compatible_candidates.size()) { break; }
const ggml_type tt = compatible_candidates[i];
const auto bpw = (float)tensor_bpw(t, tt);
const size_t bytes = tensor_bytes(t, tt);
const auto err = (float)estimate_error(t, tt, f32_sample, sample_rows_per_slice, values, activations, tl_quantized_buffer, tl_dequantised_buffer);
eval_candidates[i] = candidate_types{ tt, bpw, bytes, err };
}
});
}
for (auto &th : eval_workers) { th.join(); }
for (auto &c : eval_candidates) {
if (c.bytes > 0) { info.candidate.push_back(c); }
}
if (info.candidate.empty()) {
// As a last resort, keep original type
float bpw = ggml_nbytes(t) * 8.0f / nelem;
info.candidate.push_back(candidate_types{ t->type, bpw, ggml_nbytes(t), 0.0 });
}
// Keep only the paretooptimal candidates: if A has >= bytes and >= error than B, drop A.
{
std::vector<candidate_types> pruned;
pruned.reserve(info.candidate.size());
// Sort by bytes ascending, error ascending
std::sort(info.candidate.begin(), info.candidate.end(), [](const candidate_types & a, const candidate_types & b) {
if (a.bytes != b.bytes) { return a.bytes < b.bytes; }
return a.error < b.error;
});
double best_err = std::numeric_limits<double>::infinity();
size_t last_bytes = std::numeric_limits<size_t>::max();
for (const auto & c : info.candidate) {
// Only keep the best error seen so far at strictly larger byte sizes
if (c.bytes != last_bytes) {
// first time we see this byte size
last_bytes = c.bytes;
if (c.error < best_err) {
pruned.push_back(c);
best_err = c.error;
}
} else {
// same bytes: we already sorted by error; skip
}
}
info.candidate.swap(pruned);
}
// Initialize choice at the smallest bpw candidate
info.choice = 0;
info.min_bpw = info.candidate.front().bpw;
info.max_bpw = info.candidate.back().bpw;
all.push_back(std::move(info));
}
if (all.empty()) { return {}; }
// Greedy allocation from minimum bpw upward to reach target_bpw
auto current_total_bytes = [&]() -> size_t {
size_t b = 0;
for (const auto & ti : all) {
b += ti.candidate[ti.choice].bytes;
}
return b;
};
auto total_weights = [&]() -> size_t {
size_t w = 0;
for (const auto & ti : all) {
w += ti.n_elements;
}
return w;
};
const size_t tw = total_weights();
auto current_bpw = [&]() -> double {
return (double)current_total_bytes() * 8.0f / (double)tw;
};
// Precompute current bpw
double bpw_now = current_bpw();
float target_bpw = params->target_bpw;
// If minimal bpw is already above the target, we're constrained by the tensor's shape; return closest (min bpw)
if (bpw_now >= target_bpw) {
std::unordered_map<std::string, ggml_type> overrides;
for (const auto & ti : all) {
overrides[ggml_get_name(ti.w->tensor)] = ti.candidate[ti.choice].type;
}
return overrides;
}
struct upgrade {
int idx;
int next;
double err;
size_t delta_bytes;
double ratio;
};
// Find next strictly-larger candidate index for a tensor
auto next_distinct_idx = [&](const tensor_info & ti) -> int {
const auto & cand = ti.candidate;
const auto & cur = cand[ti.choice];
int j = ti.choice + 1;
while (j < (int)cand.size() && cand[j].bytes == cur.bytes) {
++j;
}
return j < (int)cand.size() ? j : -1;
};
auto recompute_best_upgrade = [&]() -> upgrade {
const double eps = 1e-12;
upgrade best{ -1, -1, 0.0, 0, -1.0 };
for (int i = 0; i < (int) all.size(); ++i) {
const auto & ti = all[i];
if (ti.choice >= (int)ti.candidate.size() - 1) { continue; }
const int j = next_distinct_idx(ti);
if (j < 0) { continue; }
const auto & cur = ti.candidate[ti.choice];
const auto & nxt = ti.candidate[j];
const size_t delta_bytes = nxt.bytes - cur.bytes;
if (delta_bytes == 0) { continue; }
double err = cur.error - nxt.error;
err = std::max(err, 0.0);
double ratio = err / (double)(delta_bytes * 8ull);
if (ratio > best.ratio + eps || (std::abs(ratio - best.ratio) <= eps && delta_bytes < best.delta_bytes)) {
best = upgrade{ i, j, err, delta_bytes, ratio };
}
}
return best;
};
while (true) {
upgrade up = recompute_best_upgrade();
if (up.idx < 0) { break; }
size_t now_bytes = current_total_bytes();
size_t next_bytes = now_bytes + up.delta_bytes;
double bpw_next = (double)next_bytes * 8.0 / (double)tw;
if (bpw_next <= target_bpw + 1e-12) {
all[up.idx].choice = up.next;
bpw_now = bpw_next;
} else {
break;
}
}
// We might still be below target so we try to find the best upgrade one last time
{
upgrade best_over{ -1, -1, 0.0, 0, -1.0 };
double best_over_gap = 1e300;
double under_gap = target_bpw - bpw_now;
size_t now_bytes = current_total_bytes();
for (int i = 0; i < (int) all.size(); ++i) {
const auto & ti = all[i];
if (ti.choice >= (int)ti.candidate.size() - 1) { continue; }
int j = next_distinct_idx(ti);
if (j < 0) { continue; }
const auto & cur = ti.candidate[ti.choice];
const auto & nxt = ti.candidate[j];
size_t delta_bytes = nxt.bytes - cur.bytes;
if (delta_bytes == 0) { continue; }
size_t over_bytes = now_bytes + delta_bytes;
double bpw_over = (double)over_bytes * 8.0 / (double)tw;
double err = cur.error - nxt.error;
if (err < 0.0) { err = 0.0; }
double ratio = err / (double)(delta_bytes * 8ull);
double over_gap = std::abs(bpw_over - (double)target_bpw);
if (over_gap < best_over_gap - 1e-12 || (std::abs(over_gap - best_over_gap) <= 1e-12 && ratio > best_over.ratio)) {
best_over_gap = over_gap;
best_over = upgrade{ i, j, err, delta_bytes, ratio };
}
}
if (best_over.idx >= 0) {
if (best_over_gap < under_gap) {
all[best_over.idx].choice = best_over.next;
}
}
}
// Build the override map
std::unordered_map<std::string, ggml_type> overrides;
LLAMA_LOG_INFO("%s: - estimated tensor quantization mix:\n", __func__);
for (const auto & ti : all) {
LLAMA_LOG_INFO("\t%s: %45s - \t%8s, \t%1.4f bpw,\terror: %.4f\n",
__func__, ggml_get_name(ti.w->tensor), ggml_type_name(ti.candidate[ti.choice].type), ti.candidate[ti.choice].bpw, ti.candidate[ti.choice].error);
overrides[ggml_get_name(ti.w->tensor)] = ti.candidate[ti.choice].type;
}
return overrides;
}
static void llama_model_quantize_impl(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
ggml_type default_type;
llama_ftype ftype = params->ftype;
switch (params->ftype) {
case LLAMA_FTYPE_MOSTLY_Q4_0: default_type = GGML_TYPE_Q4_0; break;
case LLAMA_FTYPE_MOSTLY_Q4_1: default_type = GGML_TYPE_Q4_1; break;
case LLAMA_FTYPE_MOSTLY_Q5_0: default_type = GGML_TYPE_Q5_0; break;
case LLAMA_FTYPE_MOSTLY_Q5_1: default_type = GGML_TYPE_Q5_1; break;
case LLAMA_FTYPE_MOSTLY_Q8_0: default_type = GGML_TYPE_Q8_0; break;
case LLAMA_FTYPE_MOSTLY_F16: default_type = GGML_TYPE_F16; break;
case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
case LLAMA_FTYPE_ALL_F32: default_type = GGML_TYPE_F32; break;
case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: default_type = GGML_TYPE_MXFP4; break;
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLAMA_FTYPE_MOSTLY_Q2_K: default_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_IQ3_XS: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLAMA_FTYPE_MOSTLY_Q3_K_M:
case LLAMA_FTYPE_MOSTLY_Q3_K_L: default_type = GGML_TYPE_Q3_K; break;
case LLAMA_FTYPE_MOSTLY_Q4_K_S:
case LLAMA_FTYPE_MOSTLY_Q4_K_M: default_type = GGML_TYPE_Q4_K; break;
case LLAMA_FTYPE_MOSTLY_Q5_K_S:
case LLAMA_FTYPE_MOSTLY_Q5_K_M: default_type = GGML_TYPE_Q5_K; break;
case LLAMA_FTYPE_MOSTLY_Q6_K: default_type = GGML_TYPE_Q6_K; break;
case LLAMA_FTYPE_MOSTLY_TQ1_0: default_type = GGML_TYPE_TQ1_0; break;
case LLAMA_FTYPE_MOSTLY_TQ2_0: default_type = GGML_TYPE_TQ2_0; break;
case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_XS: default_type = GGML_TYPE_IQ2_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_S: default_type = GGML_TYPE_IQ2_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_M: default_type = GGML_TYPE_IQ2_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ1_S: default_type = GGML_TYPE_IQ1_S; break;
case LLAMA_FTYPE_MOSTLY_IQ1_M: default_type = GGML_TYPE_IQ1_M; break;
case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break;
case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
int nthread = params->nthread;
if (nthread <= 0) {
nthread = std::thread::hardware_concurrency();
}
// mmap consistently increases speed on Linux, and also increases speed on Windows with
// hot cache. It may cause a slowdown on macOS, possibly related to free memory.
#if defined(__linux__) || defined(_WIN32)
constexpr bool use_mmap = true;
#else
constexpr bool use_mmap = false;
#endif
llama_model_kv_override * kv_overrides = nullptr;
if (params->kv_overrides) {
auto * v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
kv_overrides = v->data();
}
std::vector<std::string> splits = {};
llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr);
ml.init_mappings(false); // no prefetching
llama_model model(llama_model_default_params());
model.load_arch (ml);
model.load_hparams(ml);
model.load_stats (ml);
quantize_state_impl qs(model, params);
if (params->only_copy) {
ftype = ml.ftype;
}
const std::unordered_map<std::string, std::vector<float>> * values_data = nullptr;
const std::unordered_map<std::string, std::vector<float>> * activations_data = nullptr;
if (params->imatrix) {
values_data = static_cast<const std::unordered_map<std::string, std::vector<float>>*>(params->imatrix);
if (values_data) {
LLAMA_LOG_INFO("================================ Have weights data with %d entries",int(values_data->size()));
qs.has_imatrix = true;
// check imatrix for nans or infs
for (const auto & kv : *values_data) {
for (float f : kv.second) {
if (!std::isfinite(f)) {
throw std::runtime_error(format("imatrix contains non-finite value %f\n", f));
}
}
}
}
}
if (params->activations) {
activations_data = static_cast<const std::unordered_map<std::string, std::vector<float>>*>(params->activations);
if (activations_data) {
LLAMA_LOG_INFO(" and %d activations",int(activations_data->size()));
qs.has_activations = true;
// check activations for nans or infs
for (const auto & kv : *activations_data) {
for (float f : kv.second) {
if (!std::isfinite(f)) {
throw std::runtime_error(format("activations contain non-finite value %f\n", f));
}
}
}
}
}
LLAMA_LOG_INFO("\n");
gguf_context_ptr ctx_out { gguf_init_empty() };
std::vector<int> prune_list = {};
if (params->prune_layers) {
prune_list = *static_cast<const std::vector<int> *>(params->prune_layers);
}
// copy the KV pairs from the input file
gguf_set_kv (ctx_out.get(), ml.meta.get());
gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
gguf_set_val_u32(ctx_out.get(), "general.file_type", ftype); // TODO: use LLM_KV
// Remove split metadata
gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str());
gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str());
gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str());
if (params->kv_overrides) {
const std::vector<llama_model_kv_override> & overrides = *(const std::vector<llama_model_kv_override> *)params->kv_overrides;
for (const auto & o : overrides) {
if (o.key[0] == 0) break;
if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
} else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
// Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
} else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
} else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
gguf_set_val_str(ctx_out.get(), o.key, o.val_str);
} else {
LLAMA_LOG_WARN("%s: unknown KV override type for key %s\n", __func__, o.key);
}
}
}
std::map<int, std::string> mapped;
int blk_id = 0;
int pruned_attention_w = 0;
// make a list of weights
std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
tensors.reserve(ml.weights_map.size());
for (const auto & it : ml.weights_map) {
const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
if (remapped_name.empty()) {
if (it.first.find("attn_v.weight") != std::string::npos ||
it.first.find("attn_qkv.weight") != std::string::npos ||
it.first.find("attn_kv_b.weight") != std::string::npos) {
pruned_attention_w++;
}
LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
continue;
} else if (remapped_name != it.first) {
ggml_set_name(it.second.tensor, remapped_name.c_str());
LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
}
tensors.push_back(&it.second);
}
if (!prune_list.empty()) {
gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_BLOCK_COUNT).c_str(), blk_id);
}
// keep_split requires that the weights are sorted by split index
if (params->keep_split) {
std::sort(tensors.begin(), tensors.end(), [](const llama_model_loader::llama_tensor_weight * a, const llama_model_loader::llama_tensor_weight * b) {
if (a->idx == b->idx) {
return a->offs < b->offs;
}
return a->idx < b->idx;
});
}
for (const auto * it : tensors) {
const struct ggml_tensor * tensor = it->tensor;
const std::string name = ggml_get_name(tensor);
// TODO: avoid hardcoded tensor names - use the TN_* constants
if (name.find("attn_v.weight") != std::string::npos ||
name.find("attn_qkv.weight") != std::string::npos ||
name.find("attn_kv_b.weight")!= std::string::npos) {
++qs.n_attention_wv;
} else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
qs.has_output = true;
}
}
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
// sanity checks for models that have attention layers
if (qs.n_attention_wv != 0)
{
const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
// attention layers have a non-zero number of kv heads
int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
if (llama_model_has_encoder(&model)) {
n_attn_layer *= 3;
}
GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
}
size_t total_size_org = 0;
size_t total_size_new = 0;
std::vector<std::thread> workers;
workers.reserve(nthread);
int idx = 0;
std::vector<no_init<uint8_t>> read_data;
std::vector<no_init<uint8_t>> work;
std::vector<no_init<float>> f32_conv_buf;
uint16_t n_split = 1;
// Assume split index is continuous
if (params->keep_split) {
for (const auto * it : tensors) {
n_split = std::max(uint16_t(it->idx + 1), n_split);
}
}
std::vector<gguf_context_ptr> ctx_outs(n_split);
ctx_outs[0] = std::move(ctx_out);
// populate the original tensors so we get an initial meta data
for (const auto * it : tensors) {
uint16_t i_split = params->keep_split ? it->idx : 0;
ggml_tensor * tensor = it->tensor;
if (!ctx_outs[i_split]) {
ctx_outs[i_split].reset(gguf_init_empty());
}
gguf_add_tensor(ctx_outs[i_split].get(), tensor);
}
// Set split info if needed
if (n_split > 1) {
for (size_t i = 0; i < ctx_outs.size(); ++i) {
gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split);
gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), (int32_t)tensors.size());
}
}
std::unordered_map<std::string, ggml_type> bpw_overrides = {};
if (params->target_bpw != -1.0f && !params->only_copy) {
if (params->imatrix) {
if (params->activations) {
LLAMA_LOG_INFO("%s: imatrix with activations provided, target bpw quantization will be more accurate\n", __func__);
} else {
LLAMA_LOG_WARN("%s: imatrix without activations provided, target bpw quantization will be less accurate\n", __func__);
}
LLAMA_LOG_INFO("%s: computing tensor quantization mix to achieve %.4f bpw\n", __func__, params->target_bpw);
bpw_overrides = target_bpw_type(ml, read_data, model, tensors, mapped, values_data, activations_data, params, nthread);
} else {
LLAMA_LOG_WARN("%s: no imatrix provided, target bpw will not apply\n", __func__);
}
}
int cur_split = -1;
std::ofstream fout;
auto close_ofstream = [&]() {
// Write metadata and close file handler
if (fout.is_open()) {
fout.seekp(0);
std::vector<uint8_t> data(gguf_get_meta_size(ctx_outs[cur_split].get()));
gguf_get_meta_data(ctx_outs[cur_split].get(), data.data());
fout.write((const char *) data.data(), data.size());
fout.close();
}
};
auto new_ofstream = [&](int index) {
cur_split = index;
GGML_ASSERT(ctx_outs[cur_split] && "Find uninitialized gguf_context");
std::string fname = fname_out;
if (params->keep_split) {
std::vector<char> split_path(llama_path_max(), 0);
llama_split_path(split_path.data(), split_path.size(), fname_out.c_str(), cur_split, n_split);
fname = std::string(split_path.data());
}
fout = std::ofstream(fname, std::ios::binary);
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
const size_t meta_size = gguf_get_meta_size(ctx_outs[cur_split].get());
// placeholder for the meta data
::zeros(fout, meta_size);
};
const auto tn = LLM_TN(model.arch);
new_ofstream(0);
for (const auto * it : tensors) {
const size_t align = GGUF_DEFAULT_ALIGNMENT;
const auto & weight = *it;
ggml_tensor * tensor = weight.tensor;
if (weight.idx != cur_split && params->keep_split) {
close_ofstream();
new_ofstream(weight.idx);
}
const std::string name = ggml_get_name(tensor);
if (!ml.use_mmap) {
if (read_data.size() < ggml_nbytes(tensor)) {
read_data.resize(ggml_nbytes(tensor));
}
tensor->data = read_data.data();
}
ml.load_data_for(tensor);
LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",
++idx, ml.n_tensors, ggml_get_name(tensor), llama_format_tensor_shape(tensor).c_str(), ggml_type_name(tensor->type));
// This used to be a regex, but <regex> has an extreme cost to compile times.
bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?
// quantize only 2D and 3D tensors (experts)
quantize &= (ggml_n_dims(tensor) >= 2);
// do not quantize norm tensors
quantize &= name.find("_norm.weight") == std::string::npos;
quantize &= params->quantize_output_tensor || name != "output.weight";
quantize &= !params->only_copy;
// do not quantize expert gating tensors
// NOTE: can't use LLM_TN here because the layer number is not known
quantize &= name.find("ffn_gate_inp.weight") == std::string::npos;
// these are very small (e.g. 4x4)
quantize &= name.find("altup") == std::string::npos;
quantize &= name.find("laurel") == std::string::npos;
// these are not too big so keep them as it is
quantize &= name.find("per_layer_model_proj") == std::string::npos;
// do not quantize positional embeddings and token types (BERT)
quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD, "weight");
quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight");
// do not quantize Mamba's small yet 2D weights
// NOTE: can't use LLM_TN here because the layer number is not known
quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
quantize &= name.find("shortconv.conv.weight") == std::string::npos;
// do not quantize RWKV's small yet 2D weights
quantize &= name.find("time_mix_first.weight") == std::string::npos;
quantize &= name.find("time_mix_w0.weight") == std::string::npos;
quantize &= name.find("time_mix_w1.weight") == std::string::npos;
quantize &= name.find("time_mix_w2.weight") == std::string::npos;
quantize &= name.find("time_mix_v0.weight") == std::string::npos;
quantize &= name.find("time_mix_v1.weight") == std::string::npos;
quantize &= name.find("time_mix_v2.weight") == std::string::npos;
quantize &= name.find("time_mix_a0.weight") == std::string::npos;
quantize &= name.find("time_mix_a1.weight") == std::string::npos;
quantize &= name.find("time_mix_a2.weight") == std::string::npos;
quantize &= name.find("time_mix_g1.weight") == std::string::npos;
quantize &= name.find("time_mix_g2.weight") == std::string::npos;
quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
quantize &= name.find("time_mix_lerp_fused.weight") == std::string::npos;
// do not quantize relative position bias (T5)
quantize &= name.find("attn_rel_b.weight") == std::string::npos;
ggml_type new_type;
void * new_data;
size_t new_size;
if (quantize) {
new_type = default_type;
// get more optimal quantization type based on the tensor shape, layer, etc.
if (!params->pure && ggml_is_quantized(default_type)) {
int fallback = qs.n_fallback;
new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
// get quantization type overrides targeting a given bits per weight budget
if (params->target_bpw != -1.0f && !bpw_overrides.empty()) {
const auto override = bpw_overrides.find(name);
if (override != bpw_overrides.end() && override->second != new_type) {
LLAMA_LOG_DEBUG("(bpw override %s) ", ggml_type_name(new_type));
new_type = override->second;
}
}
// unless the user specifies a type, and the tensor shape will not require fallback quantisation
if (params->tensor_types && qs.n_fallback - fallback == 0) {
const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
const std::string tensor_name(tensor->name);
for (const auto & [tname, qtype] : tensor_types) {
if (std::regex pattern(tname); std::regex_search(tensor_name, pattern)) {
if (qtype != new_type) {
LLAMA_LOG_DEBUG("(type override %s) ", ggml_type_name(new_type));
new_type = qtype; // if two or more types are specified for the same tensor, the last match wins
}
}
}
}
}
if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
new_type = params->token_embedding_type;
}
if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) {
new_type = params->output_tensor_type;
}
// If we've decided to quantize to the same type the tensor is already
// in then there's nothing to do.
quantize = tensor->type != new_type;
}
if (!quantize) {
new_type = tensor->type;
new_data = tensor->data;
new_size = ggml_nbytes(tensor);
LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
} else {
const int64_t nelements = ggml_nelements(tensor);
const float * imatrix = nullptr;
if (values_data) {
auto it = values_data->find(remap_imatrix(tensor->name, mapped));
if (it == values_data->end()) {
LLAMA_LOG_INFO("\n====== %s: did not find weights for %s, ", __func__, tensor->name);
} else {
if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
imatrix = it->second.data();
} else {
LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
// this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
// this is a significant error and it may be good idea to abort the process if this happens,
// since many people will miss the error and not realize that most of the model is being quantized without an imatrix
// tok_embd should be ignored in this case, since it always causes this warning
if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
}
}
}
}
if ((new_type == GGML_TYPE_IQ2_XXS ||
new_type == GGML_TYPE_IQ2_XS ||
new_type == GGML_TYPE_IQ2_S ||
new_type == GGML_TYPE_IQ1_S ||
(new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight")) ||
(new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
LLAMA_LOG_ERROR("\n\n============================================================\n");
LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
LLAMA_LOG_ERROR("============================================================\n\n");
throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
}
float * f32_data;
if (tensor->type == GGML_TYPE_F32) {
f32_data = (float *) tensor->data;
} else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
} else {
llama_tensor_dequantize_impl(tensor, f32_conv_buf, workers, nelements, nthread);
f32_data = (float *) f32_conv_buf.data();
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);
if (work.size() < (size_t)nelements * 4) {
work.resize(nelements * 4); // upper bound on size
}
new_data = work.data();
const int64_t n_per_row = tensor->ne[0];
const int64_t nrows = tensor->ne[1];
static const int64_t min_chunk_size = 32 * 512;
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
// quantize each expert separately since they have different importance matrices
new_size = 0;
for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
const float * f32_data_03 = f32_data + i03 * nelements_matrix;
void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
// TODO: temporary sanity check that the F16 -> MXFP4 is lossless
#if 0
if (new_type == GGML_TYPE_MXFP4) {
auto * x = f32_data_03;
//LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
std::vector<float> deq(nrows*n_per_row);
const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
qtype->to_float(new_data_03, deq.data(), deq.size());
double err = 0.0f;
for (int i = 0; i < (int) deq.size(); ++i) {
err += fabsf(deq[i] - x[i]);
//if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
if (deq[i] != x[i]) {
LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
}
}
//LLAMA_LOG_INFO("err = %f\n", err);
GGML_ASSERT(err == 0.00000);
}
#endif
}
LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
}
total_size_org += ggml_nbytes(tensor);
total_size_new += new_size;
// update the gguf meta data as we go
gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
// write tensor data + padding
fout.write((const char *) new_data, new_size);
zeros(fout, GGML_PAD(new_size, align) - new_size);
}
close_ofstream();
LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
if (qs.n_fallback > 0) {
LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n",
__func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback);
}
}
//
// interface implementation
//
llama_model_quantize_params llama_model_quantize_default_params() {
llama_model_quantize_params result = {
/*.nthread =*/ 0,
/*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
/*.output_tensor_type =*/ GGML_TYPE_COUNT,
/*.token_embedding_type =*/ GGML_TYPE_COUNT,
/*.allow_requantize =*/ false,
/*.quantize_output_tensor =*/ true,
/*.only_copy =*/ false,
/*.pure =*/ false,
/*.keep_split =*/ false,
/*.imatrix =*/ nullptr,
/*.activations =*/ nullptr,
/*.kv_overrides =*/ nullptr,
/*.tensor_type =*/ nullptr,
/*.prune_layers =*/ nullptr,
/*.target_bpw =*/ -1.0f
};
return result;
}
uint32_t llama_model_quantize(
const char * fname_inp,
const char * fname_out,
const llama_model_quantize_params * params) {
try {
llama_model_quantize_impl(fname_inp, fname_out, params);
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what());
return 1;
}
return 0;
}