File size: 6,963 Bytes
3fd1a35 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 | #!/usr/bin/env python3
"""Export official-BF16 logits for one isolated Ling-3.0-tiny token."""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import torch
from safetensors import safe_open
HIDDEN = 1536
HEADS = 16
HEAD_DIM = 128
EXPERTS = 128
EXPERTS_PER_TOKEN = 8
GROUPS = 8
SELECTED_GROUPS = 4
ROUTED_SCALE = 2.5
class Source:
def __init__(self, root: Path):
index = json.loads((root / "model.safetensors.index.json").read_text())
self.root = root
self.weight_map = index["weight_map"]
self.handles = {
shard: safe_open(root / shard, framework="pt", device="cpu")
for shard in sorted(set(self.weight_map.values()))
}
def get(self, name: str) -> torch.Tensor:
return self.handles[self.weight_map[name]].get_tensor(name)
def rms_norm(value: torch.Tensor, weight: torch.Tensor) -> torch.Tensor:
normalized = value.float() * torch.rsqrt(value.float().square().mean() + 1.0e-6)
return (weight * normalized.to(value.dtype)).to(torch.bfloat16)
def linear(value: torch.Tensor, weight: torch.Tensor) -> torch.Tensor:
return torch.mv(weight.float(), value.float()).to(torch.bfloat16)
def mlp(source: Source, prefix: str, value: torch.Tensor) -> torch.Tensor:
gate = linear(value, source.get(prefix + ".gate_proj.weight"))
up = linear(value, source.get(prefix + ".up_proj.weight"))
activated = torch.nn.functional.silu(gate.float()).to(torch.bfloat16) * up
return linear(activated, source.get(prefix + ".down_proj.weight"))
def select_route(logits: torch.Tensor, bias: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
scores = torch.sigmoid(logits.float())
routing_scores = scores + bias.float()
group_scores = routing_scores.view(GROUPS, -1).topk(2, dim=-1).values.sum(dim=-1)
selected_groups = group_scores.topk(SELECTED_GROUPS, sorted=False).indices
mask = torch.zeros(GROUPS, dtype=torch.bool)
mask[selected_groups] = True
masked = routing_scores.masked_fill(~mask[:, None].expand(GROUPS, EXPERTS // GROUPS).reshape(-1), -torch.inf)
selected = masked.topk(EXPERTS_PER_TOKEN, sorted=False).indices
weights = scores[selected]
weights = weights / weights.sum() * ROUTED_SCALE
return selected, weights
def first_kda(source: Source, prefix: str, value: torch.Tensor) -> torch.Tensor:
def conv(name: str) -> torch.Tensor:
projected = linear(value, source.get(prefix + f".{name}_proj.weight"))
weight = source.get(prefix + f".{name}_conv1d.weight")[:, 0, -1]
return torch.nn.functional.silu(projected.float() * weight.float()).to(torch.bfloat16)
query = conv("q").view(HEADS, HEAD_DIM).float()
key = conv("k").view(HEADS, HEAD_DIM).float()
projected_value = conv("v").view(HEADS, HEAD_DIM).float()
query *= torch.rsqrt(query.square().sum(-1, keepdim=True) + 1.0e-6)
key *= torch.rsqrt(key.square().sum(-1, keepdim=True) + 1.0e-6)
beta = torch.sigmoid(linear(value, source.get(prefix + ".b_proj.weight")).float())
recurrence = beta[:, None] * projected_value * (query * key).sum(-1, keepdim=True) / (HEAD_DIM**0.5)
recurrence = recurrence.to(torch.bfloat16)
normalized = rms_norm(recurrence, source.get(prefix + ".o_norm.weight"))
gate = torch.sigmoid(linear(value, source.get(prefix + ".g_proj.weight")).float()).to(torch.bfloat16)
return linear((normalized * gate.view(HEADS, HEAD_DIM)).flatten(), source.get(prefix + ".o_proj.weight"))
def first_mla(source: Source, prefix: str, value: torch.Tensor) -> torch.Tensor:
compressed = linear(value, source.get(prefix + ".kv_a_proj_with_mqa.weight"))
latent = rms_norm(compressed[:512], source.get(prefix + ".kv_a_layernorm.weight"))
key_value = linear(latent, source.get(prefix + ".kv_b_proj.weight")).view(HEADS, 256)
attention = key_value[:, 128:].contiguous()
gate = torch.sigmoid(linear(value, source.get(prefix + ".g_proj.weight")).float()).to(torch.bfloat16)
return linear((attention * gate[:, None]).flatten(), source.get(prefix + ".dense.weight"))
def main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--source", type=Path, required=True)
parser.add_argument("--token", type=int, default=34355)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--routes", type=Path)
args = parser.parse_args()
source = Source(args.source.resolve())
hidden = source.get("model.word_embeddings.weight")[args.token].contiguous()
routes: list[dict[str, object]] = []
with torch.inference_mode():
for layer in range(24):
root = f"model.layers.{layer}"
normalized = rms_norm(hidden, source.get(root + ".input_layernorm.weight"))
attention = first_mla(source, root + ".attention", normalized) if (layer + 1) % 4 == 0 else first_kda(source, root + ".attention", normalized)
hidden = (hidden + attention).to(torch.bfloat16)
normalized = rms_norm(hidden, source.get(root + ".post_attention_layernorm.weight"))
if layer == 0:
feed_forward = mlp(source, root + ".mlp", normalized)
else:
gate_prefix = root + ".mlp.gate"
router_logits = torch.mv(source.get(gate_prefix + ".weight").float(), normalized.float())
selected, weights = select_route(router_logits, source.get(gate_prefix + ".expert_bias"))
routed = torch.zeros(HIDDEN, dtype=torch.float32)
for expert, weight in zip(selected.tolist(), weights.tolist()):
routed += mlp(source, root + f".mlp.experts.{expert}", normalized).float() * weight
routed = routed.to(torch.bfloat16)
shared = mlp(source, root + ".mlp.shared_experts", normalized)
feed_forward = (routed + shared).to(torch.bfloat16)
routes.append({
"layer": layer,
"experts": selected.tolist(),
"weights": weights.tolist(),
})
hidden = (hidden + feed_forward).to(torch.bfloat16)
print(f"layer={layer} rms={hidden.float().square().mean().sqrt().item():.9f}", flush=True)
normalized = rms_norm(hidden, source.get("model.norm.weight"))
logits = linear(normalized, source.get("lm_head.weight")).float()
args.output.parent.mkdir(parents=True, exist_ok=True)
logits.numpy().tofile(args.output)
if args.routes is not None:
args.routes.write_text(json.dumps(routes, indent=2) + "\n")
top = logits.topk(10)
print(f"token={args.token} output={args.output} elements={logits.numel()}")
print("top_ids=" + ",".join(str(value) for value in top.indices.tolist()))
print("top_logits=" + ",".join(f"{value:.8f}" for value in top.values.tolist()))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|