tiny-clef-for-testing / slice_clef.py
ngxson's picture
ngxson HF Staff
tiny test slice
b80b19b verified
Raw History Blame Contribute Delete
10.8 kB
#!/usr/bin/env python3
"""
Build a tiny slice of Cloudflare/clef (clef decision model) for testing. Output is gibberish, only the shapes and code paths are real.
Adapted from the slice-model skill of llama.cpp (skills/slice-model): tensors are fetched with HTTP range requests and config.json is
patched to match, then convert_hf_to_gguf.py does the conversion. The stock script only drops layers, which cannot go below the
248320 x 5120 token embedding, so this one also narrows the hidden sizes of the trunk, of the vision encoder and of the joint head.
usage: slice_clef.py <output dir>
"""
import json, re, struct, sys
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
import requests
import torch
from huggingface_hub import hf_hub_url
from safetensors.torch import save_file
REPO = 'Cloudflare/clef'
AUX = ['tokenizer.json', 'tokenizer_config.json', 'chat_template.jinja', 'processor_config.json', 'generation_config.json', 'LICENSE']
# text: hidden, ffn, attention heads, kv heads (head_dim stays 256), linear attention key / value heads (head dims stay 128)
E, FF, H, KV, LK, LV = 32, 128, 2, 1, 2, 6
HD, LKD, LVD = 256, 128, 128
TEXT_LAYERS = [0, 1, 2, 3] # 3 linear attention + 1 full attention
# vision: hidden, ffn, heads (head_dim stays 72)
V, VFF, VH = 144, 288, 2
VISION_LAYERS = [0, 1]
MERGE = 4 # 2 x 2 patches are merged
# joint head: width, heads (head_dim 1024 / 16 = 64 -> 16), ffn. the routing and joint layer counts are kept
JW, JH, JFF = 32, 2, 64
DTYPES = {'BF16': torch.bfloat16, 'F32': torch.float32, 'F16': torch.float16}
def read_header(url):
def rng(a, b):
r = requests.get(url, headers={'Range': f'bytes={a}-{b}'}, timeout=120)
r.raise_for_status()
return r.content
n = struct.unpack('<Q', rng(0, 7))[0]
h = json.loads(rng(8, 8 + n - 1))
h.pop('__metadata__', None)
return h, 8 + n
def fetch(url, meta, base):
a, b = meta['data_offsets']
if a == b:
return torch.zeros(meta['shape'], dtype=DTYPES[meta['dtype']])
r = requests.get(url, headers={'Range': f'bytes={base + a}-{base + b - 1}'}, timeout=600)
r.raise_for_status()
assert len(r.content) == b - a
return torch.frombuffer(bytearray(r.content), dtype=DTYPES[meta['dtype']]).reshape(meta['shape'])
def renumber(name, prefix, keep):
m = re.match(re.escape(prefix) + r'\.(\d+)\.(.*)', name)
if not m:
return name
i = int(m[1])
return f'{prefix}.{keep.index(i)}.{m[2]}' if i in keep else None
def rename_main(n):
for prefix, keep in (('model.language_model.layers', TEXT_LAYERS), ('model.visual.blocks', VISION_LAYERS)):
if n.startswith(prefix + '.'):
return renumber(n, prefix, keep)
if n.startswith('mtp.'):
return None
return n
def linear_qkv_rows(t, nk, nv):
"""in_proj_qkv / conv1d: q (nk * dk), k (nk * dk), v (nv * dv) stacked on dim 0, v heads grouped by key head."""
q, k, v = torch.split(t, [nk * LKD, nk * LKD, nv * LVD], dim=0)
return torch.cat([q[:LK * LKD], k[:LK * LKD], v[:LV * LVD]], dim=0)
def narrow_text_layer(key, t, nk, nv):
if key.endswith(('input_layernorm.weight', 'post_attention_layernorm.weight')):
return t[:E]
if key.endswith(('mlp.gate_proj.weight', 'mlp.up_proj.weight')):
return t[:FF, :E]
if key.endswith('mlp.down_proj.weight'):
return t[:E, :FF]
# full attention, q_proj has the output gate: [n_head, 2 * head_dim] per head
if key.endswith('self_attn.q_proj.weight'):
return t[:H * 2 * HD, :E]
if key.endswith(('self_attn.k_proj.weight', 'self_attn.v_proj.weight')):
return t[:KV * HD, :E]
if key.endswith('self_attn.o_proj.weight'):
return t[:E, :H * HD]
if key.endswith(('self_attn.q_norm.weight', 'self_attn.k_norm.weight', 'linear_attn.norm.weight')):
return t
# linear attention (gated delta net)
if key.endswith('linear_attn.in_proj_qkv.weight'):
return linear_qkv_rows(t, nk, nv)[:, :E]
if key.endswith('linear_attn.conv1d.weight'):
return linear_qkv_rows(t, nk, nv)
if key.endswith('linear_attn.in_proj_z.weight'):
return t[:LV * LVD, :E]
if key.endswith(('linear_attn.in_proj_a.weight', 'linear_attn.in_proj_b.weight')):
return t[:LV, :E]
if key.endswith(('linear_attn.A_log', 'linear_attn.dt_bias')):
return t[:LV]
if key.endswith('linear_attn.out_proj.weight'):
return t[:E, :LV * LVD]
raise SystemExit(f'unhandled tensor {key}')
def narrow_vision_block(key, t, vo):
if key.endswith('attn.qkv.weight'):
return t.reshape(3, vo, vo)[:, :V, :V].reshape(3 * V, V)
if key.endswith('attn.qkv.bias'):
return t.reshape(3, vo)[:, :V].reshape(-1)
if key.endswith('attn.proj.weight'):
return t[:V, :V]
if key.endswith('mlp.linear_fc1.weight'):
return t[:VFF, :V]
if key.endswith('mlp.linear_fc1.bias'):
return t[:VFF]
if key.endswith('mlp.linear_fc2.weight'):
return t[:V, :VFF]
if key.endswith(('attn.proj.bias', 'mlp.linear_fc2.bias', 'norm1.weight', 'norm1.bias', 'norm2.weight', 'norm2.bias')):
return t[:V]
raise SystemExit(f'unhandled tensor {key}')
def make_narrow(text_cfg, vision_cfg):
nk, nv = text_cfg['linear_num_key_heads'], text_cfg['linear_num_value_heads']
vo = vision_cfg['hidden_size']
def narrow(name, t):
if name.startswith('model.language_model.layers.'):
return narrow_text_layer(name, t, nk, nv)
if name.startswith('model.visual.blocks.'):
return narrow_vision_block(name, t, vo)
table = {
'model.language_model.embed_tokens.weight': lambda: t[:, :E],
'lm_head.weight': lambda: t[:, :E],
'model.language_model.norm.weight': lambda: t[:E],
'model.visual.patch_embed.proj.weight': lambda: t[:V],
'model.visual.patch_embed.proj.bias': lambda: t[:V],
'model.visual.pos_embed.weight': lambda: t[:, :V],
'model.visual.merger.norm.weight': lambda: t[:V],
'model.visual.merger.norm.bias': lambda: t[:V],
# the input of linear_fc1 is MERGE patches of the vision hidden size, its output has the same size
'model.visual.merger.linear_fc1.weight': lambda: t[:MERGE * V].reshape(MERGE * V, MERGE, vo)[:, :, :V].reshape(MERGE * V, MERGE * V),
'model.visual.merger.linear_fc1.bias': lambda: t[:MERGE * V],
'model.visual.merger.linear_fc2.weight': lambda: t[:E, :MERGE * V],
'model.visual.merger.linear_fc2.bias': lambda: t[:E],
}
if name in table:
return table[name]()
raise SystemExit(f'unhandled tensor {name}')
return narrow
def make_narrow_head(head_cfg):
w = head_cfg['width']
def narrow(name, t):
if t.ndim == 0:
return t # learned scales
if name.endswith('in_proj_weight'): # nn.MultiheadAttention: q, k, v stacked
return t.reshape(3, w, w)[:, :JW, :JW].reshape(3 * JW, JW)
if name.endswith('in_proj_bias'):
return t.reshape(3, w)[:, :JW].reshape(-1)
if name.endswith('out_proj.weight'):
return t[:JW, :JW]
if name.endswith(('linear1.weight', 'feedforward.0.weight')):
return t[:JFF, :JW]
if name.endswith(('linear1.bias', 'feedforward.0.bias')):
return t[:JFF]
if name.endswith(('linear2.weight', 'feedforward.3.weight')):
return t[:JW, :JFF]
if name.endswith('_projection.weight'): # trunk hidden -> head width
return t[:JW, :E]
if name.startswith('hidden_norm.'): # trunk hidden
return t[:E]
if name == 'residual_scorer.0.weight': # 4 vectors of the head width are concatenated
return t[:JW].reshape(JW, 4, w)[:, :, :JW].reshape(JW, 4 * JW)
if name == 'residual_scorer.3.weight':
return t[:, :JW]
if name == 'residual_scorer.3.bias':
return t
if name == 'type_embedding.weight':
return t[:, :JW]
if t.ndim == 1 and t.shape[0] == w: # norms and biases of the head width
return t[:JW]
raise SystemExit(f'unhandled tensor {name} {list(t.shape)}')
return narrow
def slice_files(filenames, out_path, rename, transform):
jobs = []
n_total = 0
for filename in filenames:
url = hf_hub_url(REPO, filename)
header, base = read_header(url)
n_total += len(header)
jobs += [(url, base, n, m, rename(n)) for n, m in header.items()]
jobs = [j for j in jobs if j[4] is not None]
def one(job):
url, base, src, meta, dst = job
return dst, transform(src, fetch(url, meta, base)).contiguous().clone()
with ThreadPoolExecutor(8) as ex:
tensors = dict(ex.map(one, jobs))
out_path.parent.mkdir(parents=True, exist_ok=True)
save_file(tensors, out_path, metadata={'format': 'pt'})
print(f'{filenames}: {len(tensors)}/{n_total} tensors, {out_path.stat().st_size / 1e6:.1f} MB')
def get_json(f):
r = requests.get(hf_hub_url(REPO, f), timeout=120)
r.raise_for_status()
return r.json()
def main():
out = Path(sys.argv[1])
out.mkdir(parents=True, exist_ok=True)
cfg = get_json('config.json')
head_cfg = get_json('joint_head_config.json')
# only the shards that hold a kept tensor
weight_map = get_json('model.safetensors.index.json')['weight_map']
shards = sorted({f for n, f in weight_map.items() if rename_main(n) is not None})
slice_files(shards, out / 'model.safetensors', rename_main, make_narrow(cfg['text_config'], cfg['vision_config']))
slice_files(['joint_head.safetensors'], out / 'joint_head.safetensors', lambda n: n, make_narrow_head(head_cfg))
for f in AUX:
r = requests.get(hf_hub_url(REPO, f), timeout=120)
r.raise_for_status()
(out / f).write_bytes(r.content)
t = cfg['text_config']
t.update(hidden_size=E, intermediate_size=FF, num_attention_heads=H, num_key_value_heads=KV,
linear_num_key_heads=LK, linear_num_value_heads=LV,
num_hidden_layers=len(TEXT_LAYERS), layer_types=[t['layer_types'][i] for i in TEXT_LAYERS])
v = cfg['vision_config']
v.update(depth=len(VISION_LAYERS), hidden_size=V, intermediate_size=VFF, num_heads=VH, out_hidden_size=E,
deepstack_visual_indexes=[i for i in v['deepstack_visual_indexes'] if i in VISION_LAYERS])
(out / 'config.json').write_text(json.dumps(cfg, indent=2) + '\n')
head_cfg.update(hidden_size=E, width=JW, heads=JH, feedforward=JFF)
(out / 'joint_head_config.json').write_text(json.dumps(head_cfg, indent=2) + '\n')
main()