reports/raw_evidence/quick_soak.log

COMMAND:
["/content/Prism-llama.cpp/build/bin/llama-completion", "-m", "/content/models/Bonsai-27B-Q1/Bonsai-27B-Q1_0.gguf", "-p", "User: Write a compact technical checklist for validating a local language model runtime. Continue until the requested token budget is reached.\nAssistant:", "-n", "192", "--seed", "1234", "--temp", "0", "--ctx-size", "256", "-ngl", "999", "--lora", "/content/prism_native_q1_lora/step10_training_pipeline/artifacts/accepted/step10_adapter_final.gguf", "--threads", "2", "--threads-batch", "2", "--batch-size", "32", "--ubatch-size", "16", "--no-display-prompt", "--no-warmup", "--ignore-eos", "-no-cnv"]

STDOUT:
<think>
Here's a thinking process:

1.  **Analyze User Input:**
   - **Task:** Write a compact technical checklist for validating a local language model runtime.
   - **Constraint:** Continue until the requested token budget is reached.
   - **Implicit Constraint:** The user didn't specify a token budget. I need to handle this carefully. I should either ask for the budget or proceed with a reasonable, compact checklist while acknowledging the constraint, or just generate a comprehensive but concise checklist and stop when it naturally reaches a reasonable length, noting that I can expand if needed. However, the prompt says "Continue until the requested token budget is reached." Since no budget is given, I'll assume they want me to generate a substantial but compact checklist, and I'll explicitly note that I'm providing a complete version, or I'll just generate it and stop when it's reasonably compact. Actually, I should probably just generate a well



STDERR:
0.00.019.860 I llama_completion: llama backend init
0.00.019.875 I llama_completion: load the model and apply lora adapter, if any
0.00.019.883 I common_init_result: fitting params to device memory ...
0.00.019.884 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)
0.03.036.680 W llama_context: n_ctx_seq (256) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
0.03.481.569 I llama_completion: llama threadpool init, n_threads = 2
0.03.483.176 I 
0.03.483.278 I system_info: n_threads = 2 (n_threads_batch = 2) / 12 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
0.03.483.283 I 
0.03.483.601 I sampler seed: 1234
0.03.483.617 I sampler params: 
	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
	top_k = 20, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.000
	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
0.03.483.627 I sampler chain: logits -> logit-bias -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
0.03.483.628 I generate: n_ctx = 256, n_batch = 32, n_predict = 192, n_keep = 0
0.03.483.632 I 
0.08.890.692 I common_perf_print:    sampling time =     136.10 ms
0.08.890.694 I common_perf_print:    samplers time =      44.60 ms /   219 tokens
0.08.890.699 I common_perf_print:        load time =       2.14 ms
0.08.890.701 I common_perf_print: prompt eval time =       0.02 ms /    27 tokens (    0.00 ms per token, 1588235.29 tokens per second)
0.08.890.704 I common_perf_print:        eval time =    4707.84 ms /   191 runs   (   24.65 ms per token,    40.57 tokens per second)
0.08.890.705 I common_perf_print:       total time =    5409.13 ms /   218 tokens
0.08.890.706 I common_perf_print: unaccounted time =     565.17 ms /  10.4 %      (total - sampling - prompt eval - eval) / (total)
0.08.890.707 I common_perf_print:    graphs reused =        190