CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

IQ3_S: a much better alternative to Q3_K (#5676)

Browse Source 
* iq4_nl: squash commits for easier rebase

* Basics (quantize, dequantize)
* CUDA dequantize and dot product
* Slightly faster CUDA dot product (120 t/s)
* Switch to 6-bit scales
* Scalar dot product
* AVX2 dot product
* ARM_NEON dot product
* Works on metal, but still slow
* Slightly better Metal dot product
* Another small Metal improvement
* Metal dot product is getting there
* Faster CUDA dot product
* Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided
* Report the actual bpw
* Add _xs mix that is 4.05 bpw for non-MoE models
* Remove IQ4_XS for now, slightly adjust kvalues_iq4nl
* AVX2 dot product uses Q8_0 instead of Q8_K
* Add to test-backend-ops
* Minor fix
* Also use use Q5_K for attn_output in MoE models
* Fixes after merging latest master
* Switching to blocks of 32
* AVX2 for blocks of 32
* Scaler dot product for blocks of 32
* ARM_NEON dot product for blocks of 32
* Metal kernels for blocks of 32
* Slightly faster Metal kernels

* Resurrecting iq3_xs

After all the experimentation, nothing was better than this.

* Minor PPL improvement via a block scale fudge factor

* Minor improvement via 3 neighbours

* iq3_xs: working scalar and AVX2 dot products

* iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s)

* iq3_xs: working Metal implementation

* Adding IQ3_M - IQ3_XS mix with mostly Q4_K

* iiq3_xs: a 3.4375 bpw variant

* iq3_xs: make CUDA work for new version

* iq3_xs: make scalar and AVX2 work for new version

* iq3_s: make ARM_NEON work with new version

* iq3_xs: make new version work on metal

Performance is very similar to Q3_K_S

* iq3_xs: tiny Metal speed improvement

* iq3_xs: tiny Metal speed improvement

* Fix stupid warning

* Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS

* iq3_xs: rename to iq3_s

* iq3_s: make tests pass

* Move Q3_K_XS mix to 3.25 bpw

* Attempt to fix failing tests

* Another attempt to fix the Windows builds

* Attempt to fix ROCm

* ROCm again

* iq3_s: partial fix for QK_K = 64

* iq3_s: make it work on metal for QK_K = 64

Pleasent surprise: the coding was super-block size independent,
so all it took was to delete some QK_K == 256 guards.

* Will this fix ROCm?

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-02-24 16:23:52 +02:00
committed by GitHub
parent 525213d2f5
commit 4c4cb30736
12 changed files with 1211 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
llama.cpp
View File

@@ -2545,6 +2545,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break;
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2890,6 +2891,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ1_S :return "IQ1_S - 1.5625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
default: return "unknown, may not work";
}
@@ -10544,6 +10547,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
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_Q3_K : GGML_TYPE_IQ3_XXS;
}
else if (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;
}
@@ -10575,13 +10584,17 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
new_type = GGML_TYPE_Q2_K;
new_type = GGML_TYPE_IQ3_XXS;
}
} else if (name.find("attn_q.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
new_type = GGML_TYPE_IQ3_XXS;
}
} 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 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
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) {
@@ -10592,6 +10605,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
: 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;
}
@@ -10623,37 +10640,41 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
if (qs.model.hparams.n_expert == 8) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_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_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
new_type = GGML_TYPE_Q5_K;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
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_Q3_K;
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_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) new_type = GGML_TYPE_Q4_K;
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_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
new_type = GGML_TYPE_Q2_K;
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_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_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
new_type = GGML_TYPE_Q2_K;
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
new_type = GGML_TYPE_IQ3_XXS;
}
++qs.i_ffn_up;
}
@@ -10673,7 +10694,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K ||
new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS ||
new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || new_type == GGML_TYPE_IQ3_S) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
@@ -10688,6 +10709,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K: new_type = GGML_TYPE_IQ4_NL; break;
@@ -10719,7 +10741,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_XS:
case LLAMA_FTYPE_MOSTLY_Q3_K_XS: quantized_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: quantized_type = GGML_TYPE_Q3_K; break;
@@ -10733,6 +10755,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ3_XXS: quantized_type = GGML_TYPE_IQ3_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ1_S: quantized_type = GGML_TYPE_IQ1_S; break;
case LLAMA_FTYPE_MOSTLY_IQ4_NL: quantized_type = GGML_TYPE_IQ4_NL; break;
case LLAMA_FTYPE_MOSTLY_IQ3_S: quantized_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_M: quantized_type = GGML_TYPE_IQ3_S; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}