Creating a minimal model loadable by llama.cpp

[MadsRC]

As a side-project, I'm attempting to create a minimal GGUF model that can successfully be loaded by llama.cpp (through llama-cpp-python) - very much related to this question: #5038

The code that I'd like to successfully run is:

import sys
from llama_cpp import Llama
Llama(model_path=sys.argv[1:][0])

or, somewhat similar from llama.cpp directly with ./llama.cpp/main -m minimal.gguf -I

I do not expect actual inference to work, I'd just like for it to be able to load the model.

Looking at gguf-py/gguf-py/examples/writer.py, I modified it as follows:

#!/usr/bin/env python3
import sys
from pathlib import Path

import numpy as np

# Necessary to load the local gguf package
#sys.path.insert(0, str(Path(__file__).parent.parent))

from gguf import GGUFWriter  # noqa: E402


# Example usage:
def writer_example() -> None:
    # Example usage with a file
    gguf_writer = GGUFWriter("example.gguf", "llama")

    gguf_writer.add_name("py007_tinyllama-1.1b-chat-v0.3")
    gguf_writer.add_context_length(2048)
    gguf_writer.add_embedding_length(2048)
    gguf_writer.add_block_count(22)
    gguf_writer.add_feed_forward_length(5632)
    gguf_writer.add_rope_dimension_count(64)
    gguf_writer.add_head_count(32)
    gguf_writer.add_head_count_kv(4)
    gguf_writer.add_layer_norm_rms_eps(0.000010)
    gguf_writer.add_rope_freq_base(10000.000000)
    gguf_writer.add_file_type(10)
    gguf_writer.add_tokenizer_model("llama")
    gguf_writer.add_token_list(["<unk>", "<s>", "</s>", "<0x00>"])
    gguf_writer.add_token_scores([0.000000, 0.000000, 0.000000, 0.000000])
    gguf_writer.add_token_types([2, 3, 3, 6])
    gguf_writer.add_bos_token_id(1)
    gguf_writer.add_eos_token_id(2)
    gguf_writer.add_unk_token_id(0)
    gguf_writer.add_quantization_version(2)

    tensor1 = np.ones((32,), dtype=np.float32) * 100.0
    tensor2 = np.ones((64,), dtype=np.float32) * 101.0
    tensor3 = np.ones((96,), dtype=np.float32) * 102.0

    gguf_writer.add_tensor("tensor1", tensor1)
    gguf_writer.add_tensor("tensor2", tensor2)
    gguf_writer.add_tensor("tensor3", tensor3)

    gguf_writer.write_header_to_file()
    gguf_writer.write_kv_data_to_file()
    gguf_writer.write_tensors_to_file()

    gguf_writer.close()


if __name__ == '__main__':
    writer_example()

That code produces a GGUF file, that if dumped with llama.cpp/gguf-py/scripts/gguf-dump.py looks like this:

* Loading: example.gguf
* File is LITTLE endian, script is running on a LITTLE endian host.

* Dumping 23 key/value pair(s)
      1: UINT32     |        1 | GGUF.version = 3
      2: UINT64     |        1 | GGUF.tensor_count = 3
      3: UINT64     |        1 | GGUF.kv_count = 20
      4: STRING     |        1 | general.architecture = 'llama'
      5: STRING     |        1 | general.name = 'py007_tinyllama-1.1b-chat-v0.3'
      6: UINT32     |        1 | llama.context_length = 2048
      7: UINT32     |        1 | llama.embedding_length = 2048
      8: UINT32     |        1 | llama.block_count = 22
      9: UINT32     |        1 | llama.feed_forward_length = 5632
     10: UINT32     |        1 | llama.rope.dimension_count = 64
     11: UINT32     |        1 | llama.attention.head_count = 32
     12: UINT32     |        1 | llama.attention.head_count_kv = 4
     13: FLOAT32    |        1 | llama.attention.layer_norm_rms_epsilon = 9.999999747378752e-06
     14: FLOAT32    |        1 | llama.rope.freq_base = 10000.0
     15: UINT32     |        1 | general.file_type = 10
     16: STRING     |        1 | tokenizer.ggml.model = 'llama'
     17: [STRING]   |        4 | tokenizer.ggml.tokens
     18: [FLOAT32]  |        4 | tokenizer.ggml.scores
     19: [INT32]    |        4 | tokenizer.ggml.token_type
     20: UINT32     |        1 | tokenizer.ggml.bos_token_id = 1
     21: UINT32     |        1 | tokenizer.ggml.eos_token_id = 2
     22: UINT32     |        1 | tokenizer.ggml.unknown_token_id = 0
     23: UINT32     |        1 | general.quantization_version = 2

* Dumping 3 tensor(s)
      1:         32 |    32,     1,     1,     1 | F32     | tensor1
      2:         64 |    64,     1,     1,     1 | F32     | tensor2
      3:         96 |    96,     1,     1,     1 | F32     | tensor3

(I straight-up took all KV values from TheBlokes tinyllm model)

My issue is that when I try to load that model up with llama.cpp, I get this:

./llama.cpp/main -m example.gguf -i
Log start
main: build = 1925 (381ee19)
main: built with Apple clang version 15.0.0 (clang-1500.0.40.1) for arm64-apple-darwin23.1.0
main: seed  = 1705703041
llama_model_loader: loaded meta data with 20 key-value pairs and 3 tensors from example.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.name str              = py007_tinyllama-1.1b-chat-v0.3
llama_model_loader: - kv   2:                       llama.context_length u32              = 2048
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 2048
llama_model_loader: - kv   4:                          llama.block_count u32              = 22
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 5632
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 64
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 4
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  11:                          general.file_type u32              = 10
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,4]       = ["<unk>", "<s>", "</s>", "<0x00>"]
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,4]       = [0.000000, 0.000000, 0.000000, 0.000000]
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,4]       = [2, 3, 3, 6]
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  19:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:    3 tensors
llama_model_load: error loading model: unordered_map::at: key not found
llama_load_model_from_file: failed to load model
llama_init_from_gpt_params: error: failed to load model 'example.gguf'
main: error: unable to load model

I can't seem to locate the source of the error loading model: unordered_map::at: key not found message, perhaps because I haven't touched C++ in a really, really long time.

Does anyone have any pointers as to how to progress on this goal? Is there a better way to build a barebones minimal model?

Full disclosure, details on the internals of actual models elude me - this here is me embarking on the journey to learn more about these internals ;)

[cmp-nct]

Not directly related to your synthetic minimal model, but maybe of interest:
https://huggingface.co/jploski/falcon-mini-shakespeare/tree/main
That's a super tiny Falcon model that can be converted and also inference, the model is less than 40MB in size.

Regarding your error message:
My first assumption is that you are missing actual content for a llama architecture.
If you have all kv from a tiny model that works, you'll still lack the vocab and actual model tensors.

You'd have to single step for the error, I'd start at init_mapping() in llama.cpp.

[Propfend]

how did you convert this model from .bin to .gguf? I didnt find anything about that on llama.cpp conversions section.

[ggerganov]

It's probably throwing when trying to find the newline token:

https://github.com/ggerganov/llama.cpp/blob/504dc37be8446fb09b1ede70300250ad41be32a2/llama.cpp#L3050-L3059

[Propfend]

I am currently having this issue:

llama-server -m example.gguf --slots --port 8080

build: 5004 (f52d59d7) with cc (Debian 12.2.0-14) 12.2.0 for x86_64-linux-gnu
system info: n_threads = 6, n_threads_batch = 6, total_threads = 12

system_info: n_threads = 6 (n_threads_batch = 6) / 12 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: binding port with default address family
main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 11
main: loading model
srv    load_model: loading model 'example.gguf'
llama_model_loader: loaded meta data with 20 key-value pairs and 3 tensors from example.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.name str              = py007_tinyllama-1.1b-chat-v0.3
llama_model_loader: - kv   2:                       llama.context_length u32              = 2048
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 2048
llama_model_loader: - kv   4:                          llama.block_count u32              = 22
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 5632
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 64
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 4
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  11:                          general.file_type u32              = 10
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,4]       = ["<unk>", "<s>", "</s>", "<0x00>"]
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,4]       = [0.000000, 0.000000, 0.000000, 0.000000]
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,4]       = [2, 3, 3, 6]
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  19:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:    3 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 0.00 MiB (32.00 BPW) 
load: SPM vocabulary, but newline token not found: unordered_map::at! Using special_pad_id instead.load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 3
load: token to piece cache size = 0.0000 MB
print_info: arch             = llama
print_info: vocab_only       = 0
print_info: n_ctx_train      = 2048
print_info: n_embd           = 2048
print_info: n_layer          = 22
print_info: n_head           = 32
print_info: n_head_kv        = 4
print_info: n_rot            = 64
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 64
print_info: n_embd_head_v    = 64
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 256
print_info: n_embd_v_gqa     = 256
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-05
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 5632
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = linear
print_info: freq_base_train  = 10000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 2048
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 1B
print_info: model params     = 0.19 K
print_info: general.name     = py007_tinyllama-1.1b-chat-v0.3
print_info: vocab type       = SPM
print_info: n_vocab          = 4
print_info: n_merges         = 0
print_info: BOS token        = 1 '<s>'
print_info: EOS token        = 2 '</s>'
print_info: UNK token        = 0 '<unk>'
print_info: EOG token        = 2 '</s>'
print_info: max token length = 6
load_tensors: loading model tensors, this can take a while... (mmap = true)
llama_model_load: error loading model: missing tensor 'token_embd.weight'
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model 'example.gguf'
srv    load_model: failed to load model, 'example.gguf'
srv    operator(): operator(): cleaning up before exit...
main: exiting due to model loading error
terminate called without an active exception
Aborted

Seems to be a problem with getting k quant type.