diff options
| author | Jared Van Bortel <jared@nomic.ai> | 2024-04-09 13:44:08 -0400 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-04-09 13:44:08 -0400 |
| commit | 1b67731e184e27a465b8c5476061294a4af668ea (patch) | |
| tree | 15a2d877029fb509a34e462c227475bc7d6dc31e /examples/imatrix | |
| parent | c4a3a4ff47d62d2503ddf9bd91b58c21f04fe3c3 (diff) | |
BERT tokenizer fixes (#6498)
Key changes:
* BERT conversion: fix abuse of LlamaHfVocab, do not set BOS or EOS
* Nomic Embed conversion: pad vocab instead of slicing embedding tensor
* llama_tokenize: handle added special tokens like HF does
Diffstat (limited to 'examples/imatrix')
| -rw-r--r-- | examples/imatrix/imatrix.cpp | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index d8cb0a64..1bf55f90 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -349,12 +349,13 @@ static void process_logits( static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl, int from_chunk) { const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1); const int n_ctx = llama_n_ctx(ctx); auto tim1 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); - std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true); auto tim2 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count()); |