granite3-moe

The IBM Granite 1B and 3B models are the first mixture of experts (MoE) Granite models from IBM designed for low latency usage.

tools 1b 3b

15.7K Pulls Updated 5 days ago

33 Tags

granite3-moe:3b-instruct-q8_0 ... /

model

441d5c119511 · 3.6GB

Metadata

general.architecture

granitemoe
general.basename

granite-3.0
general.file_type

7
general.finetune

instruct
general.license

apache-2.0
general.name

Granite 3.0 3b A800M Instruct
general.quantization_version

2
general.size_label

3B-a800M
general.tags

[language, granite-3.0, text-generation]
general.type

model
granitemoe.attention.head_count

24
granitemoe.attention.head_count_kv

8
granitemoe.attention.layer_norm_rms_epsilon

1e-06
granitemoe.attention.scale

0.015625
granitemoe.block_count

32
granitemoe.context_length

4096
granitemoe.embedding_length

1536
granitemoe.embedding_scale

12
granitemoe.expert_count

40
granitemoe.expert_used_count

8
granitemoe.feed_forward_length

512
granitemoe.logit_scale

6
granitemoe.residual_scale

0.22
granitemoe.rope.dimension_count

64
granitemoe.rope.freq_base

10000
granitemoe.vocab_size

49155
tokenizer.ggml.add_bos_token

false
tokenizer.ggml.add_space_prefix

false
tokenizer.ggml.bos_token_id

0
tokenizer.ggml.eos_token_id

0
tokenizer.ggml.merges

[Ġ Ġ, ĠĠ ĠĠ, ĠĠĠĠ ĠĠĠĠ, ĠĠ Ġ, e r, ...]
tokenizer.ggml.model

gpt2
tokenizer.ggml.padding_token_id

0
tokenizer.ggml.pre

refact
tokenizer.ggml.token_type

[3, 3, 3, 3, 3, ...]
tokenizer.ggml.tokens

[<|end_of_text|>, <fim_prefix>, <fim_middle>, <fim_suffix>, <fim_pad>, ...]

Tensor

Name

Type

Shape
token_embd.weight

Q8_0

[1536, 49155]

blk.0

blk.0.attn_k.weight

Q8_0

[1536, 512]
blk.0.attn_norm.weight

F32

[1536]
blk.0.attn_output.weight

Q8_0

[1536, 1536]
blk.0.attn_q.weight

Q8_0

[1536, 1536]
blk.0.attn_v.weight

Q8_0

[1536, 512]
blk.0.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.0.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.0.ffn_gate_inp.weight

F32

[1536, 40]
blk.0.ffn_norm.weight

F32

[1536]
blk.0.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.1

blk.1.attn_k.weight

Q8_0

[1536, 512]
blk.1.attn_norm.weight

F32

[1536]
blk.1.attn_output.weight

Q8_0

[1536, 1536]
blk.1.attn_q.weight

Q8_0

[1536, 1536]
blk.1.attn_v.weight

Q8_0

[1536, 512]
blk.1.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.1.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.1.ffn_gate_inp.weight

F32

[1536, 40]
blk.1.ffn_norm.weight

F32

[1536]
blk.1.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.2

blk.2.attn_k.weight

Q8_0

[1536, 512]
blk.2.attn_norm.weight

F32

[1536]
blk.2.attn_output.weight

Q8_0

[1536, 1536]
blk.2.attn_q.weight

Q8_0

[1536, 1536]
blk.2.attn_v.weight

Q8_0

[1536, 512]
blk.2.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.2.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.2.ffn_gate_inp.weight

F32

[1536, 40]
blk.2.ffn_norm.weight

F32

[1536]
blk.2.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.3

blk.3.attn_k.weight

Q8_0

[1536, 512]
blk.3.attn_norm.weight

F32

[1536]
blk.3.attn_output.weight

Q8_0

[1536, 1536]
blk.3.attn_q.weight

Q8_0

[1536, 1536]
blk.3.attn_v.weight

Q8_0

[1536, 512]
blk.3.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.3.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.3.ffn_gate_inp.weight

F32

[1536, 40]
blk.3.ffn_norm.weight

F32

[1536]
blk.3.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.4

blk.4.attn_k.weight

Q8_0

[1536, 512]
blk.4.attn_norm.weight

F32

[1536]
blk.4.attn_output.weight

Q8_0

[1536, 1536]
blk.4.attn_q.weight

Q8_0

[1536, 1536]
blk.4.attn_v.weight

Q8_0

[1536, 512]
blk.4.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.4.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.4.ffn_gate_inp.weight

F32

[1536, 40]
blk.4.ffn_norm.weight

F32

[1536]
blk.4.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.5

blk.5.attn_k.weight

Q8_0

[1536, 512]
blk.5.attn_norm.weight

F32

[1536]
blk.5.attn_output.weight

Q8_0

[1536, 1536]
blk.5.attn_q.weight

Q8_0

[1536, 1536]
blk.5.attn_v.weight

Q8_0

[1536, 512]
blk.5.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.5.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.5.ffn_gate_inp.weight

F32

[1536, 40]
blk.5.ffn_norm.weight

F32

[1536]
blk.5.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.6

blk.6.attn_k.weight

Q8_0

[1536, 512]
blk.6.attn_norm.weight

F32

[1536]
blk.6.attn_output.weight

Q8_0

[1536, 1536]
blk.6.attn_q.weight

Q8_0

[1536, 1536]
blk.6.attn_v.weight

Q8_0

[1536, 512]
blk.6.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.6.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.6.ffn_gate_inp.weight

F32

[1536, 40]
blk.6.ffn_norm.weight

F32

[1536]
blk.6.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.7

blk.7.attn_k.weight

Q8_0

[1536, 512]
blk.7.attn_norm.weight

F32

[1536]
blk.7.attn_output.weight

Q8_0

[1536, 1536]
blk.7.attn_q.weight

Q8_0

[1536, 1536]
blk.7.attn_v.weight

Q8_0

[1536, 512]
blk.7.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.7.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.7.ffn_gate_inp.weight

F32

[1536, 40]
blk.7.ffn_norm.weight

F32

[1536]
blk.7.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.8

blk.8.attn_k.weight

Q8_0

[1536, 512]
blk.8.attn_norm.weight

F32

[1536]
blk.8.attn_output.weight

Q8_0

[1536, 1536]
blk.8.attn_q.weight

Q8_0

[1536, 1536]
blk.8.attn_v.weight

Q8_0

[1536, 512]
blk.8.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.8.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.8.ffn_gate_inp.weight

F32

[1536, 40]
blk.8.ffn_norm.weight

F32

[1536]
blk.8.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.9

blk.9.attn_k.weight

Q8_0

[1536, 512]
blk.9.attn_norm.weight

F32

[1536]
blk.9.attn_output.weight

Q8_0

[1536, 1536]
blk.9.attn_q.weight

Q8_0

[1536, 1536]
blk.9.attn_v.weight

Q8_0

[1536, 512]
blk.9.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.9.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.9.ffn_gate_inp.weight

F32

[1536, 40]
blk.9.ffn_norm.weight

F32

[1536]
blk.9.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.10

blk.10.attn_k.weight

Q8_0

[1536, 512]
blk.10.attn_norm.weight

F32

[1536]
blk.10.attn_output.weight

Q8_0

[1536, 1536]
blk.10.attn_q.weight

Q8_0

[1536, 1536]
blk.10.attn_v.weight

Q8_0

[1536, 512]
blk.10.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.10.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.10.ffn_gate_inp.weight

F32

[1536, 40]
blk.10.ffn_norm.weight

F32

[1536]
blk.10.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.11

blk.11.attn_k.weight

Q8_0

[1536, 512]
blk.11.attn_norm.weight

F32

[1536]
blk.11.attn_output.weight

Q8_0

[1536, 1536]
blk.11.attn_q.weight

Q8_0

[1536, 1536]
blk.11.attn_v.weight

Q8_0

[1536, 512]
blk.11.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.11.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.11.ffn_gate_inp.weight

F32

[1536, 40]
blk.11.ffn_norm.weight

F32

[1536]
blk.11.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.12

blk.12.attn_k.weight

Q8_0

[1536, 512]
blk.12.attn_norm.weight

F32

[1536]
blk.12.attn_output.weight

Q8_0

[1536, 1536]
blk.12.attn_q.weight

Q8_0

[1536, 1536]
blk.12.attn_v.weight

Q8_0

[1536, 512]
blk.12.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.12.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.12.ffn_gate_inp.weight

F32

[1536, 40]
blk.12.ffn_norm.weight

F32

[1536]
blk.12.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.13

blk.13.attn_k.weight

Q8_0

[1536, 512]
blk.13.attn_norm.weight

F32

[1536]
blk.13.attn_output.weight

Q8_0

[1536, 1536]
blk.13.attn_q.weight

Q8_0

[1536, 1536]
blk.13.attn_v.weight

Q8_0

[1536, 512]
blk.13.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.13.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.13.ffn_gate_inp.weight

F32

[1536, 40]
blk.13.ffn_norm.weight

F32

[1536]
blk.13.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.14

blk.14.attn_k.weight

Q8_0

[1536, 512]
blk.14.attn_norm.weight

F32

[1536]
blk.14.attn_output.weight

Q8_0

[1536, 1536]
blk.14.attn_q.weight

Q8_0

[1536, 1536]
blk.14.attn_v.weight

Q8_0

[1536, 512]
blk.14.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.14.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.14.ffn_gate_inp.weight

F32

[1536, 40]
blk.14.ffn_norm.weight

F32

[1536]
blk.14.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.15

blk.15.attn_k.weight

Q8_0

[1536, 512]
blk.15.attn_norm.weight

F32

[1536]
blk.15.attn_output.weight

Q8_0

[1536, 1536]
blk.15.attn_q.weight

Q8_0

[1536, 1536]
blk.15.attn_v.weight

Q8_0

[1536, 512]
blk.15.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.15.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.15.ffn_gate_inp.weight

F32

[1536, 40]
blk.15.ffn_norm.weight

F32

[1536]
blk.15.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.16

blk.16.attn_k.weight

Q8_0

[1536, 512]
blk.16.attn_norm.weight

F32

[1536]
blk.16.attn_output.weight

Q8_0

[1536, 1536]
blk.16.attn_q.weight

Q8_0

[1536, 1536]
blk.16.attn_v.weight

Q8_0

[1536, 512]
blk.16.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.16.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.16.ffn_gate_inp.weight

F32

[1536, 40]
blk.16.ffn_norm.weight

F32

[1536]
blk.16.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.17

blk.17.attn_k.weight

Q8_0

[1536, 512]
blk.17.attn_norm.weight

F32

[1536]
blk.17.attn_output.weight

Q8_0

[1536, 1536]
blk.17.attn_q.weight

Q8_0

[1536, 1536]
blk.17.attn_v.weight

Q8_0

[1536, 512]
blk.17.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.17.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.17.ffn_gate_inp.weight

F32

[1536, 40]
blk.17.ffn_norm.weight

F32

[1536]
blk.17.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.18

blk.18.attn_k.weight

Q8_0

[1536, 512]
blk.18.attn_norm.weight

F32

[1536]
blk.18.attn_output.weight

Q8_0

[1536, 1536]
blk.18.attn_q.weight

Q8_0

[1536, 1536]
blk.18.attn_v.weight

Q8_0

[1536, 512]
blk.18.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.18.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.18.ffn_gate_inp.weight

F32

[1536, 40]
blk.18.ffn_norm.weight

F32

[1536]
blk.18.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.19

blk.19.attn_k.weight

Q8_0

[1536, 512]
blk.19.attn_norm.weight

F32

[1536]
blk.19.attn_output.weight

Q8_0

[1536, 1536]
blk.19.attn_q.weight

Q8_0

[1536, 1536]
blk.19.attn_v.weight

Q8_0

[1536, 512]
blk.19.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.19.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.19.ffn_gate_inp.weight

F32

[1536, 40]
blk.19.ffn_norm.weight

F32

[1536]
blk.19.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.20

blk.20.attn_k.weight

Q8_0

[1536, 512]
blk.20.attn_norm.weight

F32

[1536]
blk.20.attn_output.weight

Q8_0

[1536, 1536]
blk.20.attn_q.weight

Q8_0

[1536, 1536]
blk.20.attn_v.weight

Q8_0

[1536, 512]
blk.20.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.20.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.20.ffn_gate_inp.weight

F32

[1536, 40]
blk.20.ffn_norm.weight

F32

[1536]
blk.20.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.21

blk.21.attn_k.weight

Q8_0

[1536, 512]
blk.21.attn_norm.weight

F32

[1536]
blk.21.attn_output.weight

Q8_0

[1536, 1536]
blk.21.attn_q.weight

Q8_0

[1536, 1536]
blk.21.attn_v.weight

Q8_0

[1536, 512]
blk.21.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.21.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.21.ffn_gate_inp.weight

F32

[1536, 40]
blk.21.ffn_norm.weight

F32

[1536]
blk.21.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.22

blk.22.attn_k.weight

Q8_0

[1536, 512]
blk.22.attn_norm.weight

F32

[1536]
blk.22.attn_output.weight

Q8_0

[1536, 1536]
blk.22.attn_q.weight

Q8_0

[1536, 1536]
blk.22.attn_v.weight

Q8_0

[1536, 512]
blk.22.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.22.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.22.ffn_gate_inp.weight

F32

[1536, 40]
blk.22.ffn_norm.weight

F32

[1536]
blk.22.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.23

blk.23.attn_k.weight

Q8_0

[1536, 512]
blk.23.attn_norm.weight

F32

[1536]
blk.23.attn_output.weight

Q8_0

[1536, 1536]
blk.23.attn_q.weight

Q8_0

[1536, 1536]
blk.23.attn_v.weight

Q8_0

[1536, 512]
blk.23.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.23.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.23.ffn_gate_inp.weight

F32

[1536, 40]
blk.23.ffn_norm.weight

F32

[1536]
blk.23.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.24

blk.24.attn_k.weight

Q8_0

[1536, 512]
blk.24.attn_norm.weight

F32

[1536]
blk.24.attn_output.weight

Q8_0

[1536, 1536]
blk.24.attn_q.weight

Q8_0

[1536, 1536]
blk.24.attn_v.weight

Q8_0

[1536, 512]
blk.24.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.24.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.24.ffn_gate_inp.weight

F32

[1536, 40]
blk.24.ffn_norm.weight

F32

[1536]
blk.24.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.25

blk.25.attn_k.weight

Q8_0

[1536, 512]
blk.25.attn_norm.weight

F32

[1536]
blk.25.attn_output.weight

Q8_0

[1536, 1536]
blk.25.attn_q.weight

Q8_0

[1536, 1536]
blk.25.attn_v.weight

Q8_0

[1536, 512]
blk.25.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.25.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.25.ffn_gate_inp.weight

F32

[1536, 40]
blk.25.ffn_norm.weight

F32

[1536]
blk.25.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.26

blk.26.attn_k.weight

Q8_0

[1536, 512]
blk.26.attn_norm.weight

F32

[1536]
blk.26.attn_output.weight

Q8_0

[1536, 1536]
blk.26.attn_q.weight

Q8_0

[1536, 1536]
blk.26.attn_v.weight

Q8_0

[1536, 512]
blk.26.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.26.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.26.ffn_gate_inp.weight

F32

[1536, 40]
blk.26.ffn_norm.weight

F32

[1536]
blk.26.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.27

blk.27.attn_k.weight

Q8_0

[1536, 512]
blk.27.attn_norm.weight

F32

[1536]
blk.27.attn_output.weight

Q8_0

[1536, 1536]
blk.27.attn_q.weight

Q8_0

[1536, 1536]
blk.27.attn_v.weight

Q8_0

[1536, 512]
blk.27.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.27.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.27.ffn_gate_inp.weight

F32

[1536, 40]
blk.27.ffn_norm.weight

F32

[1536]
blk.27.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.28

blk.28.attn_k.weight

Q8_0

[1536, 512]
blk.28.attn_norm.weight

F32

[1536]
blk.28.attn_output.weight

Q8_0

[1536, 1536]
blk.28.attn_q.weight

Q8_0

[1536, 1536]
blk.28.attn_v.weight

Q8_0

[1536, 512]
blk.28.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.28.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.28.ffn_gate_inp.weight

F32

[1536, 40]
blk.28.ffn_norm.weight

F32

[1536]
blk.28.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.29

blk.29.attn_k.weight

Q8_0

[1536, 512]
blk.29.attn_norm.weight

F32

[1536]
blk.29.attn_output.weight

Q8_0

[1536, 1536]
blk.29.attn_q.weight

Q8_0

[1536, 1536]
blk.29.attn_v.weight

Q8_0

[1536, 512]
blk.29.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.29.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.29.ffn_gate_inp.weight

F32

[1536, 40]
blk.29.ffn_norm.weight

F32

[1536]
blk.29.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.30

blk.30.attn_k.weight

Q8_0

[1536, 512]
blk.30.attn_norm.weight

F32

[1536]
blk.30.attn_output.weight

Q8_0

[1536, 1536]
blk.30.attn_q.weight

Q8_0

[1536, 1536]
blk.30.attn_v.weight

Q8_0

[1536, 512]
blk.30.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.30.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.30.ffn_gate_inp.weight

F32

[1536, 40]
blk.30.ffn_norm.weight

F32

[1536]
blk.30.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

blk.31

blk.31.attn_k.weight

Q8_0

[1536, 512]
blk.31.attn_norm.weight

F32

[1536]
blk.31.attn_output.weight

Q8_0

[1536, 1536]
blk.31.attn_q.weight

Q8_0

[1536, 1536]
blk.31.attn_v.weight

Q8_0

[1536, 512]
blk.31.ffn_down_exps.weight

Q8_0

[512, 1536, 40]
blk.31.ffn_gate_exps.weight

Q8_0

[1536, 512, 40]
blk.31.ffn_gate_inp.weight

F32

[1536, 40]
blk.31.ffn_norm.weight

F32

[1536]
blk.31.ffn_up_exps.weight

Q8_0

[1536, 512, 40]

output.weight

Q8_0

[1536, 49155]
output_norm.weight

F32

[1536]