notes from deepseek attempt

added SAVE_ON_CRASH

small fix

sugar

add note to download weights

added .gitignore

Nexusflow_Athene-V2-Chat_493f1bbd561a5a7e3d27c4081d4ee47508bf6831.logits-and-weights

meta-llama_Llama-3.1-405B_b906e4dc842aa489c962f9db26554dcfdde901fe.logits.DEDUPLICATED

indexed deduplicated safetensors

meta-llama_Llama-3.3-70B-Instruct_6f6073b423013f6a7d4d9f39144961bfbfbc386b.logits.DEDUPLICATED

Nexusflow_Athene-V2-Chat_493f1bbd561a5a7e3d27c4081d4ee47508bf6831.logits.DEDUPLICATED

meta-llama_Llama-3.2-1B_4e20de362430cd3b72f300e6b0f18e50e7166e08.logits-and-weights.DEDUPLICATED

mention new constraint in readme

deduplication bugfixes. turns out safetensors does not support deduplicated tensor files.

tensor deduplication working now

added more input/output datas and drafted deduplicating tensor hash code. may not run have run into issues detering testing agian.

added missing inputs and outputs to run_test hook

fixed safetensors.py; moved scripts into scripts/

working on adding weights for athena-v2. one big bug remaining.

Merge branch 'main' of https://huggingface.co/datasets/baffo32/llm_logits

added a variable to set device map, placed model kws in another variable and passed as metadata

WritingSafeTensors generates sharded index

runs again

Merge branch 'main' of https://huggingface.co/datasets/baffo32/llm_logits

untested run_test.py code, committing to try on another system

weight downloading script

float64: meta-llama_Llama-3.1-405B_b906e4dc842aa489c962f9db26554dcfdde901fe

specify low_cpu_mem_usage=True to halve loading requirements

started a script to compare differing files a little

float64: Nexusflow_Athene-V2-Chat_493f1bbd561a5a7e3d27c4081d4ee47508bf6831.logits.safetensors

float64: meta-llama_Llama-3.3-70B-Instruct_6f6073b423013f6a7d4d9f39144961bfbfbc386b.logits.safetensors

logits-and-weights naming convention

small changes & meta-llama_Llama-3.2-1B_4e20de362430cd3b72f300e6b0f18e50e7166e08.logits.safetensors

weight compression script

provide to configure storing weights

store weights too

trying to increase accuracy and determinism

deepseek-ai_DeepSeek-V3_86518964eaef84e3fdd98e9861759a1384f9c29d_af3d32855e168b3f4c5a1a28788455f9fb4d42be

meta-llama_Llama-3.1-405B_b906e4dc842aa489c962f9db26554dcfdde901fe

deepseek-ai_DeepSeek-R1_a157fa3d494497a54586a333a23df6c2143e7697

changed safetensors metadata so files are deterministic

missing revision parameter

Nexusflow_Athene-V2-Chat_493f1bbd561a5a7e3d27c4081d4ee47508bf6831

meta-llama_Llama-3.3-70B-Instruct_6f6073b423013f6a7d4d9f39144961bfbfbc386b.logits.safetensors

meta-llama_Llama-3.2-1B_4e20de362430cd3b72f300e6b0f18e50e7166e08.logits.safetensors

deepseek-ai_DeepSeek-V3_86518964eaef84e3fdd98e9861759a1384f9c29d

meta-llama_Llama-3.3-70B-Instruct_6f6073b423013f6a7d4d9f39144961bfbfbc386b

meta-llama_Llama-3.2-1B_4e20de362430cd3b72f300e6b0f18e50e7166e08

initial commit