happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge 5b6f2f8374 into 312d870a89

This commit is contained in:
PikaPikachu 2026-03-24 05:35:01 +02:00 committed by GitHub
parent 312d870a89 5b6f2f8374
commit 56ba9059d6
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 390 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
convert_hf_to_gguf.py
View File

@ -117,7 +117,8 @@ class ModelBase:
small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
disable_mistral_community_chat_template: bool = False,
sentence_transformers_dense_modules: bool = False,
fuse_gate_up_exps: bool = False):
fuse_gate_up_exps: bool = False,
fuse_qkv: bool = False):
if type(self) is ModelBase or \
type(self) is TextModel or \
type(self) is MmprojModel:
@ -139,6 +140,10 @@ class ModelBase:
self.fuse_gate_up_exps = fuse_gate_up_exps
self._gate_exp_buffer: dict[int, Tensor] = {}
self._up_exp_buffer: dict[int, Tensor] = {}
self.fuse_qkv = fuse_qkv
self._q_buffer: dict[int, Tensor] = {}
self._k_buffer: dict[int, Tensor] = {}
self._v_buffer: dict[int, Tensor] = {}
self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
self.model_tensors = self.index_tensors(remote_hf_model_id=remote_hf_model_id)
self.metadata_override = metadata_override
@ -567,6 +572,33 @@ class ModelBase:
self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
return []
# Handle Q/K/V tensor fusion if enabled
if self.fuse_qkv and bid is not None:
if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_Q, bid):
self._q_buffer[bid] = data_torch
elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_K, bid):
self._k_buffer[bid] = data_torch
elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_V, bid):
self._v_buffer[bid] = data_torch
# Check if all three Q, K, V are buffered for this layer
if bid in self._q_buffer and bid in self._k_buffer and bid in self._v_buffer:
q_data = self._q_buffer.pop(bid)
k_data = self._k_buffer.pop(bid)
v_data = self._v_buffer.pop(bid)
# Q shape: (n_embd_q, n_embd), K shape: (n_embd_k, n_embd), V shape: (n_embd_v, n_embd)
# concatenate to (n_embd_q + n_embd_k + n_embd_v, n_embd)
fused_data = torch.cat([q_data, k_data, v_data], dim=0)
fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid)
logger.info(f"Fused Q, K, V into QKV for layer {bid}")
return [(fused_name, fused_data)]
# If we buffered a Q/K/V tensor, wait for the others
if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_Q, bid) or \
self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_K, bid) or \
self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_V, bid):
return []
return [(new_name, data_torch)]
def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
@ -12404,6 +12436,11 @@ def parse_args() -> argparse.Namespace:
help="Fuse gate_exps and up_exps tensors into a single gate_up_exps tensor for MoE models.",
)
parser.add_argument(
"--fuse-qkv", action="store_true",
help="Fuse separate Q, K, V weight tensors into a single QKV tensor.",
)
args = parser.parse_args()
if not args.print_supported_models and args.model is None:
parser.error("the following arguments are required: model")
@ -12542,7 +12579,8 @@ def main() -> None:
small_first_shard=args.no_tensor_first_split,
remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template,
sentence_transformers_dense_modules=args.sentence_transformers_dense_modules,
fuse_gate_up_exps=args.fuse_gate_up_exps
fuse_gate_up_exps=args.fuse_gate_up_exps,
fuse_qkv=args.fuse_qkv
)
if args.vocab_only:

3
gguf-py/gguf/constants.py
View File

@ -1369,6 +1369,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
@ -1706,6 +1707,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
@ -1784,6 +1786,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,

166
scripts/fuse_qkv_gguf.py Normal file
View File

@ -0,0 +1,166 @@
#!/usr/bin/env python3
"""Fuse Q/K/V tensors in an existing GGUF file into a single QKV tensor.
This script operates at the binary level to preserve ALL metadata (including
tokenizer) byte-for-byte from the original file.
Usage:
python scripts/fuse_qkv_gguf.py input.gguf output.gguf
"""
import sys, struct, os, re
import numpy as np
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'gguf-py'))
from gguf import GGUFReader
def align_offset(offset, alignment=32):
return (offset + alignment - 1) // alignment * alignment
def write_tensor_info(f, name, n_dims, dims, tensor_type, data_offset):
"""Write one tensor info entry in GGUF format."""
name_bytes = name.encode('utf-8')
f.write(struct.pack('<Q', len(name_bytes)))
f.write(name_bytes)
f.write(struct.pack('<I', n_dims))
for d in dims:
f.write(struct.pack('<Q', d))
f.write(struct.pack('<I', tensor_type))
f.write(struct.pack('<Q', data_offset))
def main():
if len(sys.argv) < 3:
print(f"Usage: {sys.argv[0]} input.gguf output.gguf")
sys.exit(1)
input_path = sys.argv[1]
output_path = sys.argv[2]
print(f"Reading {input_path}...")
reader = GGUFReader(input_path)
with open(input_path, 'rb') as f:
magic = f.read(4)
version = struct.unpack('<I', f.read(4))[0]
n_tensors_orig = struct.unpack('<Q', f.read(8))[0]
n_kv = struct.unpack('<Q', f.read(8))[0]
print(f" Version: {version}, Tensors: {n_tensors_orig}, KV fields: {n_kv}")
ti_start = min(t.field.offset for t in reader.tensors)
kv_data_start = 24
kv_data_end = ti_start
print(f" KV data: {kv_data_start} - {kv_data_end} ({kv_data_end - kv_data_start} bytes)")
tensor_map = {t.name: t for t in reader.tensors}
qkv_pattern = re.compile(r'^blk\.(\d+)\.attn_([qkv])\.weight$')
layer_qkv = {}
fused_names = set()
for name, t in tensor_map.items():
m = qkv_pattern.match(name)
if m:
layer = int(m.group(1))
qkv_type = m.group(2)
if layer not in layer_qkv:
layer_qkv[layer] = {}
layer_qkv[layer][qkv_type] = t
fused_names.add(name)
fuse_layers = sorted([l for l, d in layer_qkv.items() if len(d) == 3])
print(f" Fusing {len(fuse_layers)} layers: {fuse_layers[0]}-{fuse_layers[-1]}")
output_tensors = []
seen_layers = set()
for t in reader.tensors:
if t.name in fused_names:
m = qkv_pattern.match(t.name)
layer = int(m.group(1))
if layer in seen_layers:
continue
seen_layers.add(layer)
q = layer_qkv[layer]['q']
k = layer_qkv[layer]['k']
v = layer_qkv[layer]['v']
assert q.tensor_type == k.tensor_type == v.tensor_type
q_dims = [int(x) for x in q.field.parts[3]]
k_dims = [int(x) for x in k.field.parts[3]]
v_dims = [int(x) for x in v.field.parts[3]]
assert q_dims[0] == k_dims[0] == v_dims[0]
fused_ne0 = q_dims[0]
fused_ne1 = q_dims[1] + k_dims[1] + v_dims[1]
fused_name = f"blk.{layer}.attn_qkv.weight"
fused_data = np.concatenate([
np.array(q.data, copy=True),
np.array(k.data, copy=True),
np.array(v.data, copy=True),
])
print(f" Layer {layer}: Q{q_dims}+K{k_dims}+V{v_dims} -> QKV[{fused_ne0},{fused_ne1}] {fused_data.nbytes} bytes")
output_tensors.append((fused_name, 2, [fused_ne0, fused_ne1],
int(q.tensor_type), fused_data.tobytes()))
else:
dims = [int(x) for x in t.field.parts[3]]
n_dims = int(t.field.parts[2][0])
output_tensors.append((t.name, n_dims, dims,
int(t.tensor_type), bytes(t.data)))
n_tensors_new = len(output_tensors)
print(f"\n {n_tensors_orig} -> {n_tensors_new} tensors")
with open(input_path, 'rb') as f:
f.seek(kv_data_start)
kv_data_bytes = f.read(kv_data_end - kv_data_start)
print(f"\nWriting {output_path}...")
alignment = 32
with open(output_path, 'wb') as f:
f.write(magic)
f.write(struct.pack('<I', version))
f.write(struct.pack('<Q', n_tensors_new))
f.write(struct.pack('<Q', n_kv))
f.write(kv_data_bytes)
data_offsets = []
current_data_offset = 0
for name, n_dims, dims, ttype, data in output_tensors:
aligned = align_offset(current_data_offset, alignment)
data_offsets.append(aligned)
current_data_offset = aligned + len(data)
for i, (name, n_dims, dims, ttype, data) in enumerate(output_tensors):
write_tensor_info(f, name, n_dims, dims, ttype, data_offsets[i])
ti_section_end = f.tell()
tensor_data_start = align_offset(ti_section_end, alignment)
if tensor_data_start > ti_section_end:
f.write(b'\x00' * (tensor_data_start - ti_section_end))
for i, (name, n_dims, dims, ttype, data) in enumerate(output_tensors):
current_pos = f.tell() - tensor_data_start
target_pos = data_offsets[i]
if target_pos > current_pos:
f.write(b'\x00' * (target_pos - current_pos))
f.write(data)
final_size = f.tell()
print(f" Output size: {final_size / 1e9:.2f} GB")
print(" Done!")
if __name__ == '__main__':
main()

3
src/llama-arch.cpp
View File

@ -553,6 +553,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
@ -760,6 +761,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
@ -961,6 +963,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_OUTPUT,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,

45
src/llama-model.cpp
View File

@ -2714,6 +2714,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", bid), {n_embd_, n_ff_, n_expert_}, flags);
}
};
// helper: try merged QKV first, fall back to separate Q, K, V
auto create_tensor_qkv = [&](llama_layer & layer, int bid, int64_t n_embd_, int64_t n_embd_head_k_, int64_t n_head_, int64_t n_embd_k_gqa_, int64_t n_embd_v_gqa_) {
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_head_k_ * n_head_ + n_embd_k_gqa_ + n_embd_v_gqa_}, TENSOR_NOT_REQUIRED);
if (layer.wqkv == nullptr) {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", bid), {n_embd_, n_embd_head_k_ * n_head_}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", bid), {n_embd_, n_embd_k_gqa_}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", bid), {n_embd_, n_embd_v_gqa_}, 0);
}
};
switch (arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_REFACT:
@ -2739,9 +2749,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
// only LLaMA-family archs have fused QKV inference graph support
if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_LLAMA_EMBED) {
create_tensor_qkv(layer, i, n_embd, n_embd_head_k, n_head, n_embd_k_gqa, n_embd_v_gqa);
} else {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
}
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
// optional bias tensors
@ -3562,12 +3577,17 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
// only Qwen2 arch has fused QKV inference graph support
if (arch == LLM_ARCH_QWEN2) {
create_tensor_qkv(layer, i, n_embd, n_embd_head_k, n_head, n_embd_k_gqa, n_embd_v_gqa);
} else {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
}
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
// optional bias tensors
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
@ -3651,9 +3671,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
// only Qwen3 arch has fused QKV inference graph support
if (arch == LLM_ARCH_QWEN3) {
create_tensor_qkv(layer, i, n_embd, n_embd_head_k, n_head, n_embd_gqa, n_embd_gqa);
} else {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
}
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);

74
src/models/llama.cpp
View File

@ -43,27 +43,67 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
ggml_tensor * Qcur = nullptr;
ggml_tensor * Kcur = nullptr;
ggml_tensor * Vcur = nullptr;
if (model.layers[il].wqkv) {
// fused QKV path: one matmul, then split via views
cur = build_lora_mm(model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
const int64_t n_embd_q = n_embd_head * n_head;
const int64_t n_embd_kgq = n_embd_head * n_head_kv;
const size_t es = ggml_element_size(cur);
if (model.layers[il].bq || model.layers[il].bk || model.layers[il].bv) {
// Models with bias: view_2d -> add bias -> reshape_3d
// (ggml_add produces contiguous output, enabling reshape)
ggml_tensor * Qcur_2d = ggml_view_2d(ctx0, cur, n_embd_q, n_tokens, cur->nb[1], 0);
ggml_tensor * Kcur_2d = ggml_view_2d(ctx0, cur, n_embd_kgq, n_tokens, cur->nb[1], es * n_embd_q);
ggml_tensor * Vcur_2d = ggml_view_2d(ctx0, cur, n_embd_kgq, n_tokens, cur->nb[1], es * (n_embd_q + n_embd_kgq));
Qcur_2d = model.layers[il].bq ? ggml_add(ctx0, Qcur_2d, model.layers[il].bq) : ggml_cont(ctx0, Qcur_2d);
Kcur_2d = model.layers[il].bk ? ggml_add(ctx0, Kcur_2d, model.layers[il].bk) : ggml_cont(ctx0, Kcur_2d);
Vcur_2d = model.layers[il].bv ? ggml_add(ctx0, Vcur_2d, model.layers[il].bv) : ggml_cont(ctx0, Vcur_2d);
Qcur = ggml_reshape_3d(ctx0, Qcur_2d, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur_2d, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur_2d, n_embd_head, n_head_kv, n_tokens);
} else {
// Models without bias: view_3d directly (zero-copy, no ggml_cont needed)
Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, es * n_embd_head, cur->nb[1], 0);
Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, es * n_embd_head, cur->nb[1], es * n_embd_q);
Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, es * n_embd_head, cur->nb[1], es * (n_embd_q + n_embd_kgq));
}
cb(Qcur, "Qcur", il);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
} else {
// separate Q/K/V path
Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors,

73
src/models/qwen2.cpp
View File

@ -30,30 +30,63 @@ llm_build_qwen2::llm_build_qwen2(const llama_model & model, const llm_graph_para
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
ggml_tensor * Qcur = nullptr;
ggml_tensor * Kcur = nullptr;
ggml_tensor * Vcur = nullptr;
if (model.layers[il].wqkv) {
// fused QKV path: one matmul, then split via views
cur = build_lora_mm(model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
const int64_t n_embd_q = n_embd_head * n_head;
const int64_t n_embd_kgq = n_embd_head * n_head_kv;
const size_t es = ggml_element_size(cur);
// Models with bias: view_2d -> add bias -> reshape_3d
// (ggml_add produces contiguous output, enabling reshape)
ggml_tensor * Qcur_2d = ggml_view_2d(ctx0, cur, n_embd_q, n_tokens, cur->nb[1], 0);
ggml_tensor * Kcur_2d = ggml_view_2d(ctx0, cur, n_embd_kgq, n_tokens, cur->nb[1], es * n_embd_q);
ggml_tensor * Vcur_2d = ggml_view_2d(ctx0, cur, n_embd_kgq, n_tokens, cur->nb[1], es * (n_embd_q + n_embd_kgq));
Qcur_2d = model.layers[il].bq ? ggml_add(ctx0, Qcur_2d, model.layers[il].bq) : ggml_cont(ctx0, Qcur_2d);
Kcur_2d = model.layers[il].bk ? ggml_add(ctx0, Kcur_2d, model.layers[il].bk) : ggml_cont(ctx0, Kcur_2d);
Vcur_2d = model.layers[il].bv ? ggml_add(ctx0, Vcur_2d, model.layers[il].bv) : ggml_cont(ctx0, Vcur_2d);
Qcur = ggml_reshape_3d(ctx0, Qcur_2d, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur_2d, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur_2d, n_embd_head, n_head_kv, n_tokens);
cb(Qcur, "Qcur", il);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
} else {
// separate Q/K/V path
Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
}
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,

42
src/models/qwen3.cpp
View File

@ -30,18 +30,42 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
ggml_tensor * Qcur = nullptr;
ggml_tensor * Kcur = nullptr;
ggml_tensor * Vcur = nullptr;
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
if (model.layers[il].wqkv) {
// fused QKV path: one matmul, then split via views
cur = build_lora_mm(model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
const int64_t n_embd_q = n_embd_head * n_head;
const int64_t n_embd_kgq = n_embd_head * n_head_kv;
const size_t es = ggml_element_size(cur);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
// No bias in Qwen3: view_3d directly (zero-copy)
Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, es * n_embd_head, cur->nb[1], 0);
Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, es * n_embd_head, cur->nb[1], es * n_embd_q);
Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, es * n_embd_head, cur->nb[1], es * (n_embd_q + n_embd_kgq));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
} else {
// separate Q/K/V path
Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
}
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);