marfrit/rk-llama.cpp
1
0
Fork 0
Code Issues Pull Requests Actions 34 Packages Projects Releases Wiki Activity

model: Add EXAONE 4.5 implementations (#21733)

Browse Source 
* Add EXAONE 4.5 and Add GQA for MMproj

* mtmd: EXAONE 4.5 vision markers and projector path

EXAONE 4.5 uses <vision> and </vision> for image boundaries; Qwen keeps
<|vision_start|> and <|vision_end|>.

Route EXAONE 4.5 through the Qwen2.5-VL-style encode path (window attention
pattern, optional mmproj input norm). Update exaone4_5 projector weights and
convert_hf_to_gguf for mmproj export.

* mtmd: load EXAONE4 nextn tensors correctly

Align EXAONE4 tensor registration with EXAONE_MOE for NextN/MTP slots and avoid skip-flag propagation on duplicated rope_freqs so model loading succeeds for EXAONE 4.5 GGUF.

* Minor fixes

* Address PR feedback

* Address PR feedback

* Fix EXAONE after merge

* Fix EXAONE 4.5 conversion

* Address PR feedback

* Refactor EXAONE 4.5 conversion

* Address PR feedback

* Fix unintended deletion

* Minor fix

---------

Co-authored-by: LG-AI-EXAONE <exaonemodels@lgresearch.ai>
This commit is contained in:
Junwon Hwang
2026-06-01 18:48:53 +09:00
committed by GitHub
parent 19620004f5
commit 48b88c3b00
11 changed files with 356 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
conversion/__init__.py
+2
View File
@@ -58,6 +58,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
"Ernie4_5_ForCausalLM": "ernie",
"Ernie4_5_MoeForCausalLM": "ernie",
"EuroBertModel": "bert",
"Exaone4_5_ForConditionalGeneration": "exaone",
"Exaone4ForCausalLM": "exaone",
"ExaoneForCausalLM": "exaone",
"ExaoneMoEForCausalLM": "exaone",
@@ -240,6 +241,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
"DeepseekOCR2ForCausalLM": "deepseek",
"DeepseekOCRForCausalLM": "deepseek",
"DotsOCRForCausalLM": "dotsocr",
"Exaone4_5_ForConditionalGeneration": "exaone",
"Gemma3ForConditionalGeneration": "gemma",
"Gemma3nForConditionalGeneration": "gemma",
"Gemma4ForConditionalGeneration": "gemma",
conversion/base.py
+1 -1
View File
@@ -2593,7 +2593,7 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
# Step3-VL keeps text config under text_config but uses a custom top-level architecture.
# For text conversion we route to a dedicated text-only class.
# TODO: refactor this later to avoid adding exception here
if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM"):
if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration"):
return arch
# if "architectures" is found in the sub-config, use that instead
conversion/exaone.py
+97 -2
View File
@@ -3,14 +3,15 @@ from __future__ import annotations
import math
from pathlib import Path
from typing import Iterable, TYPE_CHECKING
from typing import Callable, Iterable, TYPE_CHECKING
import torch
if TYPE_CHECKING:
from torch import Tensor
from .base import ModelBase, TextModel, gguf
from .base import MmprojModel, ModelBase, TextModel, gguf
from .qwenvl import Qwen2VLVisionModel
@ModelBase.register("ExaoneForCausalLM")
@@ -208,3 +209,97 @@ class ExaoneMoEModel(Exaone4Model):
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("Exaone4_5_ForConditionalGeneration")
class Exaone4_5_TextModel(Exaone4Model):
"""Text tower of EXAONE 4.5; Tensors match EXAONE4"""
model_arch = gguf.MODEL_ARCH.EXAONE4
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0) or 0)
if n_nextn > 0:
self.block_count = self.hparams["num_hidden_layers"] + n_nextn
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
def set_gguf_parameters(self):
super().set_gguf_parameters()
n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0) or 0)
if n_nextn > 0:
self.gguf_writer.add_nextn_predict_layers(n_nextn)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("mtp."):
n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0) or 0)
if n_nextn <= 0:
return
nh = self.hparams["num_hidden_layers"]
if ".layers." in name:
share = self.hparams.get("mtp_share_layers", False)
mtp_bid = bid if bid is not None else 0
if share:
for k in range(n_nextn):
nn = name.replace(f"mtp.layers.{mtp_bid}", f"model.layers.{nh + k}")
yield from super().modify_tensors(data_torch, nn, nh + k)
return
name = name.replace(f"mtp.layers.{mtp_bid}", f"model.layers.{mtp_bid + nh}")
else:
remapper = {
"mtp.fc": gguf.MODEL_TENSOR.NEXTN_EH_PROJ,
"mtp.pre_fc_norm_embedding": gguf.MODEL_TENSOR.NEXTN_ENORM,
"mtp.pre_fc_norm_hidden": gguf.MODEL_TENSOR.NEXTN_HNORM,
"mtp.norm": gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
}
_n = Path(name)
key = _n.stem
if key not in remapper:
return
for bid_mtp in range(nh, self.block_count):
mapped_name = self.format_tensor_name(remapper[key], bid_mtp, suffix=_n.suffix)
yield from ModelBase.modify_tensors(self, data_torch, mapped_name, bid_mtp)
return
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("Exaone4_5_ForConditionalGeneration")
class Exaone4_5VisionModel(Qwen2VLVisionModel):
"""Vision tower for EXAONE 4.5; Qwen2-VL-style ViT (GQA) + patch merger"""
@classmethod
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
name, gen = item
name = name.replace("model.visual.", "visual.", 1)
return super().filter_tensors((name, gen))
def set_gguf_parameters(self):
MmprojModel.set_gguf_parameters(self)
assert self.hparams_vision is not None
hparams = self.hparams_vision
self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.EXAONE4_5)
self.gguf_writer.add_vision_use_silu(True)
self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
num_kv_head = self.find_vparam(["num_key_value_heads"], optional=True)
if num_kv_head is not None:
self.gguf_writer.add_vision_head_count_kv(num_kv_head)
eps = hparams.get("rms_norm_eps", self.global_config.get("rms_norm_eps", 1e-6))
self.gguf_writer.add_vision_attention_layernorm_eps(eps)
if (window_size := hparams.get("window_size")) is not None:
self.gguf_writer.add_vision_window_size(window_size)
fullatt_block_indexes = hparams.get("fullatt_block_indexes")
if fullatt_block_indexes:
n_wa_pattern = fullatt_block_indexes[0] + 1
for i in range(1, len(fullatt_block_indexes)):
if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern:
raise ValueError(f"Invalid EXAONE4.5 fullatt_block_indexes: {fullatt_block_indexes}")
self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if ".qkv." in name:
yield from ModelBase.modify_tensors(self, data_torch, name, bid)
return
yield from Qwen2VLVisionModel.modify_tensors(self, data_torch, name, bid)
gguf-py/gguf/constants.py
+8
View File
@@ -3310,6 +3310,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_POST_NORM,
# NextN/MTP tensors - preserved but unused
MODEL_TENSOR.NEXTN_EH_PROJ,
MODEL_TENSOR.NEXTN_EMBED_TOKENS,
MODEL_TENSOR.NEXTN_ENORM,
MODEL_TENSOR.NEXTN_HNORM,
MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
],
MODEL_ARCH.EXAONE_MOE: [
MODEL_TENSOR.TOKEN_EMBD,
@@ -4318,6 +4325,7 @@ class VisionProjectorType:
LLAMA4 = "llama4"
QWEN2VL = "qwen2vl_merger"
QWEN25VL = "qwen2.5vl_merger"
EXAONE4_5 = "exaone4_5"
QWEN3VL = "qwen3vl_merger"
STEP3VL = "step3vl"
ULTRAVOX = "ultravox"
src/models/exaone4.cpp
+35 -12
View File
@@ -15,6 +15,9 @@ void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
switch (hparams.n_layer) {
case 30: type = LLM_TYPE_1_2B; break;
@@ -38,21 +41,37 @@ void llama_model_exaone4::load_arch_tensors(llama_model_loader &) {
}
for (int i = 0; i < n_layer; ++i) {
const bool is_nextn = hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers;
int flags = 0;
if (is_nextn) {
// NextN/MTP layers are preserved in GGUF but are not executed yet.
flags |= TENSOR_SKIP;
}
auto & layer = layers[i];
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, flags);
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
if (!is_nextn) {
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
}
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags);
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);
if (is_nextn) {
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), {n_embd}, flags);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), {n_embd}, flags);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
}
}
}
@@ -90,7 +109,11 @@ llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_gra
}
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
// MTP / NextN tail blocks are loaded for compatibility but not executed (same as exaone-moe).
const int n_layer_main = int(n_layer) - int(hparams.nextn_predict_layers);
GGML_ASSERT(n_layer_main > 0);
for (int il = 0; il < n_layer_main; ++il) {
ggml_tensor * inpSA = inpL;
// use RoPE for SWA layers or non-SWA models
@@ -126,7 +149,7 @@ llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_gra
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
cb(cur, "attn_out", il);
}
if (il == n_layer - 1 && inp_out_ids) {
if (il == n_layer_main - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
tools/mtmd/CMakeLists.txt
+1
View File
@@ -18,6 +18,7 @@ add_library(mtmd
models/cogvlm.cpp
models/conformer.cpp
models/dotsocr.cpp
models/exaone4_5.cpp
models/gemma4a.cpp
models/gemma4v.cpp
models/glm4v.cpp
tools/mtmd/clip-impl.h
+2
View File
@@ -347,6 +347,7 @@ enum projector_type {
PROJECTOR_TYPE_KIMIK25,
PROJECTOR_TYPE_NEMOTRON_V2_VL,
PROJECTOR_TYPE_HUNYUANVL,
PROJECTOR_TYPE_EXAONE4_5,
PROJECTOR_TYPE_MINICPMV4_6,
PROJECTOR_TYPE_GRANITE_SPEECH,
PROJECTOR_TYPE_MIMOVL,
@@ -396,6 +397,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_YASA2, "yasa2"},
{ PROJECTOR_TYPE_KIMIK25, "kimik25"},
{ PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"},
{ PROJECTOR_TYPE_EXAONE4_5, "exaone4_5"},
{ PROJECTOR_TYPE_HUNYUANVL, "hunyuanvl"},
{ PROJECTOR_TYPE_MINICPMV4_6, "minicpmv4_6"},
{ PROJECTOR_TYPE_GRANITE_SPEECH, "granite_speech"},
tools/mtmd/clip.cpp
+28 -1
View File
@@ -884,6 +884,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
{
builder = std::make_unique<clip_graph_qwen3vl>(ctx, img);
} break;
case PROJECTOR_TYPE_EXAONE4_5:
{
builder = std::make_unique<clip_graph_exaone4_5>(ctx, img);
} break;
case PROJECTOR_TYPE_MIMOVL:
{
builder = std::make_unique<clip_graph_mimovl>(ctx, img);
@@ -1556,6 +1560,19 @@ struct clip_model_loader {
hparams.audio_window_len = 400;
hparams.audio_hop_len = 160;
} break;
case PROJECTOR_TYPE_EXAONE4_5:
{
hparams.n_merge = 2;
get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, false);
get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
hparams.set_warmup_n_tokens(46 * 46);
if (hparams.rope_theta <= 0.0f) {
hparams.rope_theta = 10000.0f;
}
get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
} break;
case PROJECTOR_TYPE_GEMMA4A:
{
// Gemma4 feature_extraction_gemma4.py:
@@ -1799,6 +1816,7 @@ struct clip_model_loader {
|| model.proj_type == PROJECTOR_TYPE_LDPV2
|| model.proj_type == PROJECTOR_TYPE_QWEN2VL
|| model.proj_type == PROJECTOR_TYPE_QWEN25VL
|| model.proj_type == PROJECTOR_TYPE_EXAONE4_5
|| model.proj_type == PROJECTOR_TYPE_GLM_EDGE
|| model.proj_type == PROJECTOR_TYPE_GEMMA3
|| model.proj_type == PROJECTOR_TYPE_IDEFICS3
@@ -1948,6 +1966,7 @@ struct clip_model_loader {
} break;
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_EXAONE4_5:
{
model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
@@ -3079,6 +3098,7 @@ int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 *
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_EXAONE4_5:
case PROJECTOR_TYPE_MIMOVL:
case PROJECTOR_TYPE_GLM4V:
case PROJECTOR_TYPE_PADDLEOCR:
@@ -3100,6 +3120,7 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_EXAONE4_5:
case PROJECTOR_TYPE_MIMOVL:
case PROJECTOR_TYPE_GLM4V:
case PROJECTOR_TYPE_PADDLEOCR:
@@ -3179,6 +3200,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_EXAONE4_5:
case PROJECTOR_TYPE_MIMOVL:
case PROJECTOR_TYPE_GLM4V:
case PROJECTOR_TYPE_YOUTUVL:
@@ -3675,11 +3697,15 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
set_input_i32("positions", positions);
} break;
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_EXAONE4_5:
case PROJECTOR_TYPE_YOUTUVL:
{
// pw * ph = number of tokens output by ViT after apply patch merger
// ipw * ipw = number of vision token been processed inside ViT
const bool use_window_attn = ctx->model.proj_type == PROJECTOR_TYPE_QWEN25VL ? hparams.n_wa_pattern > 0 : !hparams.wa_layer_indexes.empty();
const bool use_window_attn =
(ctx->model.proj_type == PROJECTOR_TYPE_QWEN25VL || ctx->model.proj_type == PROJECTOR_TYPE_EXAONE4_5)
? hparams.n_wa_pattern > 0
: !hparams.wa_layer_indexes.empty();
const int merge_ratio = 2;
const int pw = image_size_width / patch_size / merge_ratio;
const int ph = image_size_height / patch_size / merge_ratio;
@@ -4262,6 +4288,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return ctx->model.mm_model_mlp_3_w->ne[1];
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_EXAONE4_5:
case PROJECTOR_TYPE_JANUS_PRO:
case PROJECTOR_TYPE_YOUTUVL:
return ctx->model.mm_1_b->ne[0];
tools/mtmd/models/exaone4_5.cpp
+170
View File
@@ -0,0 +1,170 @@
// similar to qwen2vl, except for GQA attention
#include "models.h"
ggml_cgraph * clip_graph_exaone4_5::build() {
GGML_ASSERT(model.patch_bias == nullptr);
GGML_ASSERT(model.class_embedding == nullptr);
const int batch_size = 1;
const bool use_window_attn = hparams.n_wa_pattern > 0;
const int n_wa_pattern = hparams.n_wa_pattern;
const int n_pos = n_patches;
const int num_position_ids = n_pos * 4;
const norm_type norm_t = NORM_TYPE_RMS;
const int64_t n_kv_head = hparams.n_head_kv > 0 ? hparams.n_head_kv : n_head;
GGML_ASSERT(n_head % n_kv_head == 0);
int rope_sections[4] = { d_head / 4, d_head / 4, d_head / 4, d_head / 4 };
const float rope_freq_base = hparams.rope_theta > 0.0f ? hparams.rope_theta : 10000.0f;
ggml_tensor * inp_raw = build_inp_raw();
ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
GGML_ASSERT(img.nx % (patch_size * 2) == 0);
GGML_ASSERT(img.ny % (patch_size * 2) == 0);
{
ggml_tensor * inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
inp = ggml_add(ctx0, inp, inp_1);
inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);
inp = ggml_cont_4d(
ctx0, inp,
n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
inp = ggml_reshape_4d(
ctx0, inp,
n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
inp = ggml_cont_3d(
ctx0, inp,
n_embd, n_patches_x * n_patches_y, batch_size);
}
ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
ggml_set_name(positions, "positions");
ggml_set_input(positions);
ggml_tensor * window_mask = nullptr;
ggml_tensor * window_idx = nullptr;
ggml_tensor * inv_window_idx = nullptr;
if (use_window_attn) {
window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
ggml_set_name(window_idx, "window_idx");
ggml_set_input(window_idx);
inv_window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4);
ggml_set_name(inv_window_idx, "inv_window_idx");
ggml_set_input(inv_window_idx);
window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
ggml_set_name(window_mask, "window_mask");
ggml_set_input(window_mask);
if (flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
window_mask = ggml_cast(ctx0, window_mask, GGML_TYPE_F16);
}
}
ggml_tensor * inpL = inp;
if (use_window_attn) {
GGML_ASSERT(batch_size == 1);
inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);
inpL = ggml_get_rows(ctx0, inpL, inv_window_idx);
inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_patches_x * n_patches_y, batch_size);
}
for (int il = 0; il < n_layer; il++) {
const auto & layer = model.layers[il];
const bool full_attn = use_window_attn ? (il + 1) % n_wa_pattern == 0 : true;
ggml_tensor * cur = inpL;
cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
cb(cur, "ln1", il);
{
GGML_ASSERT(layer.qkv_w != nullptr);
cur = build_mm(layer.qkv_w, cur);
if (layer.qkv_b) {
cur = ggml_add(ctx0, cur, layer.qkv_b);
}
const int64_t n_embd_kv = d_head * n_kv_head;
ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_patches,
ggml_row_size(cur->type, d_head),
cur->nb[1],
0);
ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_kv_head, n_patches,
ggml_row_size(cur->type, d_head),
cur->nb[1],
ggml_row_size(cur->type, n_embd));
ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_kv_head, n_patches,
ggml_row_size(cur->type, d_head),
cur->nb[1],
ggml_row_size(cur->type, n_embd + n_embd_kv));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_multi(
ctx0, Qcur, positions, nullptr,
d_head / 2, rope_sections, GGML_ROPE_TYPE_VISION, 32768, rope_freq_base, 1, 0, 1, 32, 1);
Kcur = ggml_rope_multi(
ctx0, Kcur, positions, nullptr,
d_head / 2, rope_sections, GGML_ROPE_TYPE_VISION, 32768, rope_freq_base, 1, 0, 1, 32, 1);
cb(Qcur, "Qcur_rope", il);
cb(Kcur, "Kcur_rope", il);
cb(Vcur, "Vcur", il);
ggml_tensor * attn_mask = full_attn ? nullptr : window_mask;
cur = build_attn(layer.o_w, layer.o_b, Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
cb(cur, "attn_out", il);
}
cur = ggml_add(ctx0, cur, inpL);
inpL = cur;
cb(cur, "ffn_inp", il);
cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il);
cb(cur, "ffn_inp_normed", il);
cur = build_ffn(cur,
layer.ff_up_w, layer.ff_up_b,
layer.ff_gate_w, layer.ff_gate_b,
layer.ff_down_w, layer.ff_down_b,
hparams.ffn_op, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, inpL, cur);
cb(cur, "layer_out", il);
inpL = cur;
}
ggml_tensor * embeddings = inpL;
embeddings = build_norm(embeddings, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer);
embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
embeddings = build_ffn(embeddings,
model.mm_0_w, model.mm_0_b,
nullptr, nullptr,
model.mm_1_w, model.mm_1_b,
FFN_GELU,
-1);
if (use_window_attn) {
GGML_ASSERT(batch_size == 1);
embeddings = ggml_reshape_2d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4);
embeddings = ggml_get_rows(ctx0, embeddings, window_idx);
embeddings = ggml_reshape_3d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4, batch_size);
}
ggml_build_forward_expand(gf, embeddings);
return gf;
}
tools/mtmd/models/models.h
+5
View File
@@ -196,3 +196,8 @@ struct clip_graph_kimik25 : clip_graph {
ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
};
struct clip_graph_exaone4_5 : clip_graph {
clip_graph_exaone4_5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
};
tools/mtmd/mtmd.cpp
+7
View File
@@ -505,6 +505,13 @@ struct mtmd_context {
img_end = "<｜hy_place▁holder▁no▁101｜>";
image_preproc = std::make_unique<mtmd_image_preprocessor_dyn_size>(ctx_v);
} break;
case PROJECTOR_TYPE_EXAONE4_5:
{
// <vision> ... (image embeddings) ... </vision>
img_beg = "<vision>";
img_end = "</vision>";
image_preproc = std::make_unique<mtmd_image_preprocessor_dyn_size>(ctx_v);
} break;
default:
throw std::runtime_error(string_format("%s: unexpected vision projector type %d\n", __func__, proj));
}