text stringlengths 0 663 |
|---|
[2026-06-08 21:38:47] |
=== GGUF Model Evaluation === |
[2026-06-08 21:38:47] Timestamp : 20260608_213847 |
[2026-06-08 21:38:47] Eval dir : /tmp/llama_eval/evaluation |
[2026-06-08 21:38:47] Temp dir : /tmp/working/quants_for_model_base |
[2026-06-08 21:38:47] DAT dir : /tmp/llama_dat |
[2026-06-08 21:38:47] ❌ ERROR: llama-perplexity not found in any known location. |
Searched: |
/content/llama_bins/cuda-12.8 |
/kaggle/working/llama_bins |
/content/llama_bins |
/workspace/llama_bin |
/workspace/llama_bins |
FIX: Run build_llama_cpp_gpu.py first, then re-run this script. |
If already built, pass --llama-bin-dir pointing to the folder containing llama-perplexity. |
[2026-06-08 21:38:47] Traceback (most recent call last): |
File "/kaggle/working/MoQ/Model_Evaluation/evaluate_models.py", line 493, in main |
binary_dir, llama_perplexity = resolve_binary_dir(args.llama_bin_dir) |
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
File "/kaggle/working/MoQ/Model_Evaluation/evaluate_models.py", line 107, in resolve_binary_dir |
raise RuntimeError( |
RuntimeError: llama-perplexity not found in any known location. |
Searched: |
/content/llama_bins/cuda-12.8 |
/kaggle/working/llama_bins |
/content/llama_bins |
/workspace/llama_bin |
/workspace/llama_bins |
FIX: Run build_llama_cpp_gpu.py first, then re-run this script. |
If already built, pass --llama-bin-dir pointing to the folder containing llama-perplexity. |
[2026-06-08 21:41:03] |
=== GGUF Model Evaluation === |
[2026-06-08 21:41:03] Timestamp : 20260608_214103 |
[2026-06-08 21:41:03] Eval dir : /tmp/llama_eval/evaluation |
[2026-06-08 21:41:03] Temp dir : /tmp/working/quants_for_model_base |
[2026-06-08 21:41:03] DAT dir : /tmp/llama_dat |
[2026-06-08 21:41:03] Binary dir: /workspace/llama_bins/cuda-12.8 |
[2026-06-08 21:41:03] Binary : /workspace/llama_bins/cuda-12.8/llama-perplexity |
[2026-06-08 21:41:03] Target models to evaluate: 17 |
[2026-06-08 21:41:03] Using default test file: /tmp/wiki_test.raw |
[2026-06-08 21:41:03] Test files: 1 |
[2026-06-08 21:41:03] |
📥 Downloading base model: unsloth/gemma-4-12b-it-GGUF / gemma-4-12b-it-BF16.gguf |
[2026-06-08 21:42:41] ✅ Base model downloaded: /tmp/working/quants_for_model_base/gemma-4-12b-it-BF16.gguf |
[2026-06-08 21:42:42] |
============================================================ |
[2026-06-08 21:42:42] PHASE 1: Generating base logits for all files |
[2026-06-08 21:42:42] ============================================================ |
[2026-06-08 21:42:42] [1/1] Base model on /tmp/wiki_test.raw → /tmp/llama_dat/base_logits_0.dat |
[2026-06-08 21:42:42] Running: /workspace/llama_bins/cuda-12.8/llama-perplexity -m /tmp/working/quants_for_model_base/gemma-4-12b-it-BF16.gguf -f /tmp/wiki_test.raw -c 512 -b 512 -np 1 --kl-divergence-base /tmp/llama_dat/base_logits_0.dat --seed 1337 -ngl 99 --chunks 50 |
[2026-06-08 21:42:44] 0.00.354.937 I common_init_result: fitting params to device memory ... |
[2026-06-08 21:42:44] 0.00.354.944 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on) |
[2026-06-08 21:42:46] 0.03.273.212 W load: control-looking token: 212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden |
[2026-06-08 21:42:46] 0.03.274.221 W load: control-looking token: 50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden |
[2026-06-08 21:42:46] 0.03.287.624 W load: control-looking token: 1 '<eos>' was not control-type; this is probably a bug in the model. its type will be overridden |
[2026-06-08 21:42:47] 0.03.338.053 W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list |
[2026-06-08 21:44:48] 2.04.668.163 W llama_context: n_ctx_seq (512) < n_ctx_train (262144) -- the full capacity of the model will not be utilized |
[2026-06-08 21:44:48] 2.04.719.186 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
[2026-06-08 21:45:46] 3.03.061.386 I |
[2026-06-08 21:45:47] 3.03.065.339 I system_info: n_threads = 2 (n_threads_batch = 2) / 4 | CUDA : ARCHS = 750,800,860,890,900,1000,1200 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | |
[2026-06-08 21:45:47] 3.03.068.725 I perplexity: saving all logits to /tmp/llama_dat/base_logits_0.dat |
[2026-06-08 21:45:47] 3.03.068.754 I perplexity: tokenizing the input .. |
[2026-06-08 21:45:47] 3.04.264.133 I perplexity: tokenization took 1195.37 ms |
[2026-06-08 21:45:47] 3.04.264.658 I perplexity: calculating perplexity over 50 chunks, n_ctx=512, batch_size=512, n_seq=1 |
[2026-06-08 21:46:13] 3.29.594.631 I perplexity: 25.15 seconds per pass - ETA 20.95 minutes |
[2026-06-08 21:50:49] [1]333.8534,[2]692.5704,[3]593.3470,[4]571.4367,[5]597.1249,[6]638.0777,[7]726.1804,[8]782.8555,[9]839.7019,[10]876.1415,[11]892.0461,[12]880.8686,[13]898.3565,[14]899.9181,[15]903.4237,[16]899.4092,[17]784.5126,[18]809.6776,[19]801.5940,[20]784.7842,[21]771.8415,[22]766.9286,[23]738.6668,[24]695.... |
[2026-06-08 21:50:49] 8.05.302.638 I Final estimate: PPL = 592.2767 +/- 30.22510 |
[2026-06-08 21:50:49] |
[2026-06-08 21:50:52] |
Phase 1 complete. Valid files: 1 / 1 |
[2026-06-08 21:50:54] 🗑️ Deleted /tmp/working/quants_for_model_base/gemma-4-12b-it-BF16.gguf to free disk space |
[2026-06-08 21:50:54] ============================================================ |
[2026-06-08 21:50:54] PHASE 2: Running KL divergence for each target model |
[2026-06-08 21:50:54] ============================================================ |
[2026-06-08 21:50:54] |
============================================================ |
[2026-06-08 21:50:54] TARGET MODEL 1/17: kaitchup/gemma-4-12b-it-GGUF-MoQ / MoQ-3.0.gguf |
[2026-06-08 21:50:54] ============================================================ |
[2026-06-08 21:51:47] |
[1/1] MoQ-3.0.gguf vs /tmp/wiki_test.raw (base: /tmp/llama_dat/base_logits_0.dat) |
[2026-06-08 21:51:48] Running: /workspace/llama_bins/cuda-12.8/llama-perplexity -m /tmp/working/quants_for_model_base/MoQ-3.0.gguf -f /tmp/wiki_test.raw -c 512 -b 512 -np 1 --kl-divergence-base /tmp/llama_dat/base_logits_0.dat --kl-divergence --seed 1337 -ngl 99 --chunks 50 |
[2026-06-08 21:51:49] 0.00.373.192 I common_init_result: fitting params to device memory ... |
[2026-06-08 21:51:49] 0.00.373.202 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on) |
[2026-06-08 21:51:52] 0.02.903.899 W load: control-looking token: 212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden |
[2026-06-08 21:51:52] 0.02.904.806 W load: control-looking token: 50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden |
[2026-06-08 21:51:52] 0.02.916.194 W load: control-looking token: 1 '<eos>' was not control-type; this is probably a bug in the model. its type will be overridden |
[2026-06-08 21:51:52] 0.02.952.952 W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list |
[2026-06-08 21:51:53] 0.04.459.034 W llama_context: n_ctx_seq (512) < n_ctx_train (131072) -- the full capacity of the model will not be utilized |
[2026-06-08 21:51:53] 0.04.502.260 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
[2026-06-08 21:52:07] 0.18.282.767 I |
[2026-06-08 21:52:07] 0.18.282.949 I system_info: n_threads = 2 (n_threads_batch = 2) / 4 | CUDA : ARCHS = 750,800,860,890,900,1000,1200 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | |
[2026-06-08 21:52:07] 0.18.380.399 I kl_divergence: computing over 50 chunks, n_ctx=512, batch_size=512, n_seq=1 |
[2026-06-08 21:53:27] 1.38.320.479 I kl_divergence: 79.94 seconds per pass - ETA 1 hours 6.60 minutes |
[2026-06-08 21:53:28] |
[2026-06-08 21:53:28] chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p |
[2026-06-08 21:53:28] 1 1254.5112 ± 457.5198 1.58224 ± 0.22140 3.22998 ± 0.20012 38.721 ± 2.398 % 38.039 ± 3.046 % |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 41