{"schema":"latentlens.glassbox/1","status":"ok","model_id":"Qwen/Qwen3-8B-Base","revision":"49e3418fbbbca6ecbdf9608b4d22e5a407081db4","dtype":"bfloat16","loader":"stream","attn_implementation":"eager","text":"The researcher opened the notebook","tokens":[{"pos":0,"id":151643,"text":"<|endoftext|>","prefix":true},{"pos":1,"id":785,"text":"The","prefix":false},{"pos":2,"id":31085,"text":" researcher","prefix":false},{"pos":3,"id":8930,"text":" opened","prefix":false},{"pos":4,"id":279,"text":" the","prefix":false},{"pos":5,"id":37166,"text":" notebook","prefix":false}],"query_position":5,"n_heads":32,"n_kv_heads":8,"head_dim":128,"hidden_size":4096,"intermediate_size":12288,"rms_norm_eps":1e-06,"rope_theta":1000000,"n_layers":36,"layers":[{"layer":0,"file":"layer_00.bin","floats":104896,"arrays":[{"name":"resid_pre","shape":[4096],"offset":0},{"name":"gain_in","shape":[4096],"offset":4096},{"name":"attn_norm","shape":[4096],"offset":8192},{"name":"q_raw","shape":[32,128],"offset":12288},{"name":"gain_q","shape":[128],"offset":16384},{"name":"q_normed","shape":[32,128],"offset":16512},{"name":"q_rope","shape":[32,128],"offset":20608},{"name":"k_raw","shape":[8,128],"offset":24704},{"name":"gain_k","shape":[128],"offset":25728},{"name":"k_normed","shape":[8,128],"offset":25856},{"name":"k_rope_all","shape":[6,8,128],"offset":26880},{"name":"v_all","shape":[6,8,128],"offset":33024},{"name":"attn_probs","shape":[32,6],"offset":39168},{"name":"attn_concat","shape":[32,128],"offset":39360},{"name":"o_out","shape":[4096],"offset":43456},{"name":"resid_mid","shape":[4096],"offset":47552},{"name":"gain_post","shape":[4096],"offset":51648},{"name":"mlp_norm","shape":[4096],"offset":55744},{"name":"gate","shape":[12288],"offset":59840},{"name":"up","shape":[12288],"offset":72128},{"name":"act_prod","shape":[12288],"offset":84416},{"name":"down","shape":[4096],"offset":96704},{"name":"resid_post","shape":[4096],"offset":100800}],"python_ledger":[{"check":"RMSNorm #1 (input_layernorm)","max_err_rel_to_max":0.004649,"median_rel_err":0.001682,"uses":"resid_pre, learned gain"},{"check":"Q RMSNorm (q_norm, per head)","max_err_rel_to_max":0.003689,"median_rel_err":0.001584,"uses":"q_proj output, q_norm gain"},{"check":"K RMSNorm (k_norm, per head)","max_err_rel_to_max":0.002344,"median_rel_err":0.001138,"uses":"k_proj output, k_norm gain"},{"check":"RoPE on Q (rotate-half, theta)","max_err_rel_to_max":0.006931,"median_rel_err":0.00076,"uses":"q after q_norm, position, theta"},{"check":"scores, scaling and softmax (causal: all keys up to the query)","max_err_rel_to_max":0.013353,"median_rel_err":0.007654,"uses":"q_rope, k_rope of every position"},{"check":"weighted sum of values (per head, grouped KV)","max_err_rel_to_max":0.002216,"median_rel_err":0.00116,"uses":"attention probabilities, v of every position"},{"check":"residual addition #1","max_err_rel_to_max":0.002878,"median_rel_err":0.000952,"uses":"resid_pre, o_proj output"},{"check":"RMSNorm #2 (post_attention_layernorm)","max_err_rel_to_max":0.001893,"median_rel_err":0.001534,"uses":"resid_mid, learned gain"},{"check":"SwiGLU: SiLU(gate) x up","max_err_rel_to_max":0.001805,"median_rel_err":0.001046,"uses":"gate_proj and up_proj outputs"},{"check":"residual addition #2","max_err_rel_to_max":0.001389,"median_rel_err":0.001035,"uses":"resid_mid, down_proj output"}]},{"layer":17,"file":"layer_17.bin","floats":104896,"arrays":[{"name":"resid_pre","shape":[4096],"offset":0},{"name":"gain_in","shape":[4096],"offset":4096},{"name":"attn_norm","shape":[4096],"offset":8192},{"name":"q_raw","shape":[32,128],"offset":12288},{"name":"gain_q","shape":[128],"offset":16384},{"name":"q_normed","shape":[32,128],"offset":16512},{"name":"q_rope","shape":[32,128],"offset":20608},{"name":"k_raw","shape":[8,128],"offset":24704},{"name":"gain_k","shape":[128],"offset":25728},{"name":"k_normed","shape":[8,128],"offset":25856},{"name":"k_rope_all","shape":[6,8,128],"offset":26880},{"name":"v_all","shape":[6,8,128],"offset":33024},{"name":"attn_probs","shape":[32,6],"offset":39168},{"name":"attn_concat","shape":[32,128],"offset":39360},{"name":"o_out","shape":[4096],"offset":43456},{"name":"resid_mid","shape":[4096],"offset":47552},{"name":"gain_post","shape":[4096],"offset":51648},{"name":"mlp_norm","shape":[4096],"offset":55744},{"name":"gate","shape":[12288],"offset":59840},{"name":"up","shape":[12288],"offset":72128},{"name":"act_prod","shape":[12288],"offset":84416},{"name":"down","shape":[4096],"offset":96704},{"name":"resid_post","shape":[4096],"offset":100800}],"python_ledger":[{"check":"RMSNorm #1 (input_layernorm)","max_err_rel_to_max":0.003121,"median_rel_err":0.001472,"uses":"resid_pre, learned gain"},{"check":"Q RMSNorm (q_norm, per head)","max_err_rel_to_max":0.002858,"median_rel_err":0.001568,"uses":"q_proj output, q_norm gain"},{"check":"K RMSNorm (k_norm, per head)","max_err_rel_to_max":0.002923,"median_rel_err":0.001517,"uses":"k_proj output, k_norm gain"},{"check":"RoPE on Q (rotate-half, theta)","max_err_rel_to_max":0.003134,"median_rel_err":0.00085,"uses":"q after q_norm, position, theta"},{"check":"scores, scaling and softmax (causal: all keys up to the query)","max_err_rel_to_max":0.014838,"median_rel_err":0.009255,"uses":"q_rope, k_rope of every position"},{"check":"weighted sum of values (per head, grouped KV)","max_err_rel_to_max":0.00202,"median_rel_err":0.001318,"uses":"attention probabilities, v of every position"},{"check":"residual addition #1","max_err_rel_to_max":0.000781,"median_rel_err":0.000327,"uses":"resid_pre, o_proj output"},{"check":"RMSNorm #2 (post_attention_layernorm)","max_err_rel_to_max":0.002066,"median_rel_err":0.001571,"uses":"resid_mid, learned gain"},{"check":"SwiGLU: SiLU(gate) x up","max_err_rel_to_max":0.002757,"median_rel_err":0.001384,"uses":"gate_proj and up_proj outputs"},{"check":"residual addition #2","max_err_rel_to_max":0.001078,"median_rel_err":0.000246,"uses":"resid_mid, down_proj output"}]},{"layer":35,"file":"layer_35.bin","floats":104896,"arrays":[{"name":"resid_pre","shape":[4096],"offset":0},{"name":"gain_in","shape":[4096],"offset":4096},{"name":"attn_norm","shape":[4096],"offset":8192},{"name":"q_raw","shape":[32,128],"offset":12288},{"name":"gain_q","shape":[128],"offset":16384},{"name":"q_normed","shape":[32,128],"offset":16512},{"name":"q_rope","shape":[32,128],"offset":20608},{"name":"k_raw","shape":[8,128],"offset":24704},{"name":"gain_k","shape":[128],"offset":25728},{"name":"k_normed","shape":[8,128],"offset":25856},{"name":"k_rope_all","shape":[6,8,128],"offset":26880},{"name":"v_all","shape":[6,8,128],"offset":33024},{"name":"attn_probs","shape":[32,6],"offset":39168},{"name":"attn_concat","shape":[32,128],"offset":39360},{"name":"o_out","shape":[4096],"offset":43456},{"name":"resid_mid","shape":[4096],"offset":47552},{"name":"gain_post","shape":[4096],"offset":51648},{"name":"mlp_norm","shape":[4096],"offset":55744},{"name":"gate","shape":[12288],"offset":59840},{"name":"up","shape":[12288],"offset":72128},{"name":"act_prod","shape":[12288],"offset":84416},{"name":"down","shape":[4096],"offset":96704},{"name":"resid_post","shape":[4096],"offset":100800}],"python_ledger":[{"check":"RMSNorm #1 (input_layernorm)","max_err_rel_to_max":0.003404,"median_rel_err":0.00159,"uses":"resid_pre, learned gain"},{"check":"Q RMSNorm (q_norm, per head)","max_err_rel_to_max":0.002982,"median_rel_err":0.001572,"uses":"q_proj output, q_norm gain"},{"check":"K RMSNorm (k_norm, per head)","max_err_rel_to_max":0.002471,"median_rel_err":0.001507,"uses":"k_proj output, k_norm gain"},{"check":"RoPE on Q (rotate-half, theta)","max_err_rel_to_max":0.001418,"median_rel_err":0.000822,"uses":"q after q_norm, position, theta"},{"check":"scores, scaling and softmax (causal: all keys up to the query)","max_err_rel_to_max":0.008721,"median_rel_err":0.007206,"uses":"q_rope, k_rope of every position"},{"check":"weighted sum of values (per head, grouped KV)","max_err_rel_to_max":0.001984,"median_rel_err":0.001316,"uses":"attention probabilities, v of every position"},{"check":"residual addition #1","max_err_rel_to_max":0.00142,"median_rel_err":0.000955,"uses":"resid_pre, o_proj output"},{"check":"RMSNorm #2 (post_attention_layernorm)","max_err_rel_to_max":0.00289,"median_rel_err":0.001305,"uses":"resid_mid, learned gain"},{"check":"SwiGLU: SiLU(gate) x up","max_err_rel_to_max":0.001578,"median_rel_err":8.5e-05,"uses":"gate_proj and up_proj outputs"},{"check":"residual addition #2","max_err_rel_to_max":0.002463,"median_rel_err":0.000538,"uses":"resid_mid, down_proj output"}]}],"byte_order":"little-endian float32","notes":["All tensors are for the query position (the last token) except k_rope_all and v_all, which hold every position.","The model computed in bfloat16; values were converted to float32 for storage, so re-derivations agree to bfloat16 precision.","Matrix multiplications by q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj and the LM head need the weight matrices (about 16 GB) and are shown as recorded inputs and outputs, not re-derived."],"runtime_s":24.7,"provenance":{"git_commit":"cc886f7b88092f5940b11bfb8a063226f42288a2-dirty","timestamp":"2026-10-16T22:04:58+00:00","versions":{"python":"3.13.14","torch":"2.14.1+cu126","transformers":"5.18.0","tokenizers":"0.23.2","safetensors":"0.8.0","numpy":"2.5.3","scikit-learn":"1.9.1","pandas":"3.0.6"},"hardware":{"platform":"Windows-11-10.0.26300-SP0","machine":"AMD64","node":"BrahmaVishnuShiva","gpu":"NVIDIA GeForce RTX 3050 Ti Laptop GPU","gpu_memory_gib":4.0},"script":"scripts/latentlens/capture_glassbox.py","loader":"stream"}}