GGML Q4_0 Quantize Analysis in llama.cpp| 东毅居士

GGML Q4_0 Quantize Analysis in llama.cpp

作者：XD / 发表： 2023年11月13日 20:39 / 更新： 2023年11月13日 21:32 / 科研学习 / 阅读量：851

GGML Q4_0 Quantization in llama.cpp

For the LLAMA7B model, there are 387 tensors consisting of various weights and biases. These tensors include token_embd.weight, 32 sets of attention and feedforward network weights and biases (attn_norm.weigh, attn_q.weight, attn_k.weight, attn_v.weight, attn_q.bias, attn_k.bias, attn_v.bias, attn_output.weight, ffn_norm.weight, ffn_up.weight, ffn_gate.weight, ffn_down.weight), output_norm.weight, and output.weight.

Quantization Details:

Total Tensors for Quantization: 226
token_embd.weight
32 sets of: attn_q.weight, attn_k.weight, attn_v.weight, attn_output.weight, ffn_up.weight, ffn_gate.weight, ffn_down.weight
output.weight

Tensor Breakdown:

llama_model_loader:
f32 type: 161 tensors
f16 type: 226 tensors
llama_model_quantize_internal:
Meta size: 6162784 bytes

Example Tensors:

[ 1/ 387] token_embd.weight - [ 4096, 151851, 1, 1], type = f16, quantizing to q4_0 .. size = 1186.34 MB -> 333.66 MB | hist: 0.036 0.016 0.025 0.039 0.057 0.077 0.096 0.111 0.117 0.111 0.096 0.077 0.057 0.039 0.025 0.021

[ 2/ 387] blk.0.attn_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB

[ 3/ 387] blk.0.attn_q.weight - [ 4096, 4096, 1, 1], type = f16, quantizing to q4_0 .. size = 32.00 MB -> 9.00 MB | hist: 0.036 0.015 0.025 0.038 0.056 0.076 0.097 0.113 0.120 0.113 0.097 0.076 0.056 0.038 0.025 0.020
[ 4/ 387] blk.0.attn_k.weight - [ 4096, 4096, 1, 1], type = f16, quantizing to q4_0 .. size = 32.00 MB -> 9.00 MB | hist: 0.036 0.015 0.024 0.037 0.055 0.075 0.097 0.115 0.123 0.115 0.097 0.076 0.055 0.037 0.024 0.020
[ 5/ 387] blk.0.attn_v.weight - [ 4096, 4096, 1, 1], type = f16, quantizing to q4_0 .. size = 32.00 MB -> 9.00 MB | hist: 0.036 0.016 0.025 0.039 0.056 0.076 0.096 0.112 0.119 0.112 0.096 0.076 0.056 0.039 0.025 0.021

[ 6/ 387] blk.0.attn_q.bias - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB

[ 7/ 387] blk.0.attn_k.bias - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB

[ 8/ 387] blk.0.attn_v.bias - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB

[ 9/ 387] blk.0.attn_output.weight - [ 4096, 4096, 1, 1], type = f16, quantizing to q4_0 .. size = 32.00 MB -> 9.00 MB | hist: 0.036 0.015 0.025 0.039 0.056 0.077 0.096 0.112 0.118 0.112 0.096 0.077 0.056 0.039 0.025 0.021

[ 10/ 387] blk.0.ffn_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB

[ 11/ 387] blk.0.ffn_up.weight - [ 4096, 11008, 1, 1], type = f16, quantizing to q4_0 .. size = 86.00 MB -> 24.19 MB | hist: 0.037 0.016 0.025 0.039 0.057 0.077 0.096 0.111 0.116 0.111 0.096 0.077 0.057 0.039 0.025 0.021
[ 12/ 387] blk.0.ffn_gate.weight - [ 4096, 11008, 1, 1], type = f16, quantizing to q4_0 .. size = 86.00 MB -> 24.19 MB | hist: 0.037 0.016 0.026 0.039 0.057 0.077 0.096 0.110 0.116 0.110 0.096 0.077 0.057 0.040 0.026 0.021
[ 13/ 387] blk.0.ffn_down.weight - [11008, 4096, 1, 1], type = f16, quantizing to q4_0 .. size = 86.00 MB -> 24.19 MB | hist: 0.036 0.016 0.025 0.039 0.057 0.077 0.096 0.111 0.117 0.111 0.096 0.077 0.057 0.039 0.025 0.021

*...and so on for other tensors [14/ 387]-[385/ 387] *

The remaining 31 blocks follow a similar pattern. blk.0*-blk.31*

[ 386/ 387] output_norm.weight - [ 4096, 1, 1, 1], type = f32, size = 0.016 MB