vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

adding debug notes

Browse source
This commit is contained in:
mike dupont 2023-11-21 15:11:18 -05:00
parent 22359f7afe
commit 5432cc18da
2 changed files with 50 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

1
.gitignore vendored
View file

@ -99,3 +99,4 @@ tests/test-tokenizer-0-llama
tests/test-tokenizer-0-falcon tests/test-tokenizer-0-falcon
tests/test-tokenizer-1-llama tests/test-tokenizer-1-llama
tests/test-tokenizer-1-bpe tests/test-tokenizer-1-bpe
/#llama.cpp#

49
README.org
View file

@ -1037,3 +1037,52 @@ nm /mnt/data1/2023/11/09/llama.cpp/build/bin/main >main.nm
grep libcuda report7.gron -C10 > cudareport.txt grep libcuda report7.gron -C10 > cudareport.txt
grep -C1000 libcuda report7.jq > cuda.txt grep -C1000 libcuda report7.jq > cuda.txt
(gpt_params &) @0x7fffffffc960: {seed = 1700596789, n_threads = 12,
n_threads_batch = -1, n_predict = -1, n_ctx = 512, n_batch = 512, n_keep = 0,
n_draft = 16, n_chunks = -1, n_parallel = 1, n_sequences = 1, p_accept = 0.5,
p_split = 0.100000001, n_gpu_layers = -1, n_gpu_layers_draft = -1, main_gpu = 0,
tensor_split = {0 <repeats 16 times>}, n_beams = 0, rope_freq_base = 0,
--Type <RET> for more, q to quit, c to continue without paging--
rope_freq_scale = 0, yarn_ext_factor = -1, yarn_attn_factor = 1,
yarn_beta_fast = 32, yarn_beta_slow = 1, yarn_orig_ctx = 0,
rope_scaling_type = -1 '\377', sparams = {n_prev = 64, n_probs = 0, top_k = 40,
top_p = 0.949999988, min_p = 0.0500000007, tfs_z = 1, typical_p = 1,
temp = 0.800000012, penalty_last_n = 64, penalty_repeat = 1.10000002,
--Type <RET> for more, q to quit, c to continue without paging--
penalty_freq = 0, penalty_present = 0, mirostat = 0, mirostat_tau = 5,
mirostat_eta = 0.100000001, penalize_nl = true, grammar = "",
cfg_negative_prompt = "", cfg_scale = 1,
logit_bias = std::unordered_map with 0 elements},
model = "/home/mdupont/.ollama/models/mistral", model_draft = "",
--Type <RET> for more, q to quit, c to continue without paging--
model_alias = "unknown", prompt = "", prompt_file = "", path_prompt_cache = "",
input_prefix = "", input_suffix = "",
antiprompt = std::vector of length 0, capacity 0, logdir = "",
lora_adapter = std::vector of length 0, capacity 0, lora_base = "", ppl_stride = 0,
ppl_output_type = 0, hellaswag = false, hellaswag_tasks = 400, mul_mat_q = true,
--Type <RET> for more, q to quit, c to continue without paging--
memory_f16 = true, random_prompt = false, use_color = false, interactive = false,
chatml = false, prompt_cache_all = false, prompt_cache_ro = false,
embedding = false, escape = false, interactive_first = false,
multiline_input = false, simple_io = false, cont_batching = false,
input_prefix_bos = false, ignore_eos = false, instruct = false, logits_all = false,
--Type <RET> for more, q to quit, c to continue without paging--
use_mmap = true, use_mlock = false, numa = false, verbose_prompt = false,
infill = false, mmproj = "", image = ""}
(gdb)
llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
at /home/mdupont/experiments/llama.cpp/ggml.cpp:18561
18561 ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
p *ctx
$14 = {header = {magic = "GGUF", version = 2, n_tensors = 291, n_kv = 20},
kv = 0x555556ffc2f0, infos = 0x55555716d5f0, alignment = 0, offset = 0, size = 0,
data = 0x0}
(gdb)
l