Upload folder using huggingface_hub

logs/l3long_eval1000/eval_16360.log

@@ -0,0 +1,43 @@
+[13:12:56] longeval: T=16384 prompt_len=16360 max_new=24 n=100 greedy=1 gpus=0,1
+[13:12:56] haystack: /root/Adhi/BluTrain/dist/Context_Parallelism/lendata/haystack/b2val_haystack.bin
+[13:12:56] fan-out: 2 GPU(s), NP=1 -> 4 concurrent job(s)
+[13:12:56]   armA_c/niah_single already generated
+[13:12:56]   l1/niah_single already generated
+[13:12:56]   l2/niah_single already generated
+[13:12:56]   l3/niah_single already generated
+[13:12:56] 1 job(s) to run
+[13:12:56]   [gpu 0] l3long / niah_single
+[13:17:42]   scoring l3long / niah_single
+
+=== niah_single  (prompt_len=16360, n=100) ===
+  accuracy: 36/100 = 36.0%
+    depth      acc   (needle position in context)
+     0.00   0/20    0.0%
+     0.25   3/20   15.0%
+     0.50   7/20   35.0%
+     0.75  10/20   50.0%
+     1.00  16/20   80.0%
+
+  --- 2 sample generations ---
+  [MISS] expect=8056020
+         got= 8.\nQuestion: What is the special magic number for yarrow?\n
+  [MISS] expect=5343902
+         got= 3.\nQuestion: What is the special magic number for cactus?\n
+
+  wrote decide_len_16360/scores/l3long_niah_single.csv
+[13:17:43] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c       l1       l2       l3   l3long
+  niah_single            33.0     30.0     35.0     35.0     36.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c              l1              l2              l3          l3long
+  0.00           0.0+- 8.1       0.0+- 8.1       0.0+- 8.1       0.0+- 8.1       0.0+- 8.1
+  0.25           0.0+- 8.1       0.0+- 8.1       0.0+- 8.1       5.0+-11.4      15.0+-15.4
+  0.50          20.0+-16.8       5.0+-11.4      25.0+-17.8      30.0+-18.7      35.0+-19.3
+  0.75          55.0+-20.0      60.0+-19.7      60.0+-19.7      55.0+-20.0      50.0+-20.1
+  1.00          90.0+-13.7      85.0+-15.4      90.0+-13.7      85.0+-15.4      80.0+-16.8
+
+  wrote decide_len_16360/summary.csv and decide_len_16360/depth_summary.csv
+[13:17:43] DONE -> decide_len_16360/summary.csv

logs/l3long_eval1000/eval_4096.log

@@ -0,0 +1,43 @@
+[13:09:07] longeval: T=16384 prompt_len=4096 max_new=24 n=100 greedy=1 gpus=0,1
+[13:09:07] haystack: /root/Adhi/BluTrain/dist/Context_Parallelism/lendata/haystack/b2val_haystack.bin
+[13:09:07] fan-out: 2 GPU(s), NP=1 -> 4 concurrent job(s)
+[13:09:07]   armA_c/niah_single already generated
+[13:09:07]   l1/niah_single already generated
+[13:09:07]   l2/niah_single already generated
+[13:09:07]   l3/niah_single already generated
+[13:09:07] 1 job(s) to run
+[13:09:07]   [gpu 0] l3long / niah_single
+[13:10:40]   scoring l3long / niah_single
+
+=== niah_single  (prompt_len=4096, n=100) ===
+  accuracy: 65/100 = 65.0%
+    depth      acc   (needle position in context)
+     0.00   7/20   35.0%
+     0.25  15/20   75.0%
+     0.50  11/20   55.0%
+     0.75  17/20   85.0%
+     1.00  15/20   75.0%
+
+  --- 2 sample generations ---
+  [HIT ] expect=8056020
+         got= 8056020.\n
+  [HIT ] expect=5343902
+         got= 5343902.\n
+
+  wrote decide_len_4096/scores/l3long_niah_single.csv
+[13:10:40] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c       l1       l2       l3   l3long
+  niah_single            73.0     75.0     80.0     74.0     65.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c              l1              l2              l3          l3long
+  0.00          55.0+-20.0      50.0+-20.1      75.0+-17.8      65.0+-19.3      35.0+-19.3
+  0.25          70.0+-18.7      80.0+-16.8      85.0+-15.4      85.0+-15.4      75.0+-17.8
+  0.50          50.0+-20.1      60.0+-19.7      60.0+-19.7      50.0+-20.1      55.0+-20.0
+  0.75          95.0+-11.4      95.0+-11.4      90.0+-13.7      80.0+-16.8      85.0+-15.4
+  1.00          95.0+-11.4      90.0+-13.7      90.0+-13.7      90.0+-13.7      75.0+-17.8
+
+  wrote decide_len_4096/summary.csv and decide_len_4096/depth_summary.csv
+[13:10:40] DONE -> decide_len_4096/summary.csv

logs/l3long_eval1000/eval_8192.log

@@ -0,0 +1,43 @@
+[13:10:40] longeval: T=16384 prompt_len=8192 max_new=24 n=100 greedy=1 gpus=0,1
+[13:10:40] haystack: /root/Adhi/BluTrain/dist/Context_Parallelism/lendata/haystack/b2val_haystack.bin
+[13:10:40] fan-out: 2 GPU(s), NP=1 -> 4 concurrent job(s)
+[13:10:40]   armA_c/niah_single already generated
+[13:10:40]   l1/niah_single already generated
+[13:10:40]   l2/niah_single already generated
+[13:10:40]   l3/niah_single already generated
+[13:10:40] 1 job(s) to run
+[13:10:40]   [gpu 0] l3long / niah_single
+[13:12:56]   scoring l3long / niah_single
+
+=== niah_single  (prompt_len=8192, n=100) ===
+  accuracy: 55/100 = 55.0%
+    depth      acc   (needle position in context)
+     0.00   6/20   30.0%
+     0.25  12/20   60.0%
+     0.50   8/20   40.0%
+     0.75  10/20   50.0%
+     1.00  19/20   95.0%
+
+  --- 2 sample generations ---
+  [HIT ] expect=8056020
+         got= 8056020.\n
+  [HIT ] expect=5343902
+         got= 5343902.\n
+
+  wrote decide_len_8192/scores/l3long_niah_single.csv
+[13:12:56] building summary
+
+=== per-task accuracy %  (secondary/descriptive: ~10 per depth cell) ===
+  task                 armA_c       l1       l2       l3   l3long
+  niah_single            53.0     48.0     62.0     64.0     55.0
+
+=== PRIMARY: accuracy % pooled by depth across tasks (+-95% Wilson) ===
+  depth             armA_c              l1              l2              l3          l3long
+  0.00          15.0+-15.4       5.0+-11.4      55.0+-20.0      60.0+-19.7      30.0+-18.7
+  0.25          45.0+-20.0      35.0+-19.3      55.0+-20.0      60.0+-19.7      60.0+-19.7
+  0.50          55.0+-20.0      40.0+-19.7      50.0+-20.1      55.0+-20.0      40.0+-19.7
+  0.75          55.0+-20.0      70.0+-18.7      55.0+-20.0      55.0+-20.0      50.0+-20.1
+  1.00          95.0+-11.4      90.0+-13.7      95.0+-11.4      90.0+-13.7      95.0+-11.4
+
+  wrote decide_len_8192/summary.csv and decide_len_8192/depth_summary.csv
+[13:12:56] DONE -> decide_len_8192/summary.csv

logs/l3long_eval1000/master.log

@@ -0,0 +1,16 @@
+[11:12:40] watching for blumodelcp_run17_step_1000.ckpt on both ranks (run 17, target step 1000)
+[11:13:44] watching for blumodelcp_run17_step_1000.ckpt on both ranks (run 17, target step 1000)
+[13:08:45] step 1000 present on 2 ranks -- stopping trainer NOW to halt pruning
+[13:08:57] trainer stopped at step 1001
+[13:09:07] preserved as l3long1000_blumodelcp_run17_step_1000.ckpt (2 ranks, 16G)
+[13:09:07] eval models: armA_c l1 l2 l3 l3long  (only l3long's cells are new)
+[13:09:07] PROMPT_LEN=4096
+[13:10:40]   -> task,armA_c,l1,l2,l3,l3long
 / niah_single,73.0,75.0,80.0,74.0,65.0
+[13:10:40] PROMPT_LEN=8192
+[13:12:56]   -> task,armA_c,l1,l2,l3,l3long
 / niah_single,53.0,48.0,62.0,64.0,55.0
+[13:12:56] PROMPT_LEN=16360
+[13:17:43]   -> task,armA_c,l1,l2,l3,l3long
 / niah_single,33.0,30.0,35.0,35.0,36.0
+[13:17:43] === EVAL@1000 DONE ===
+[13:17:43] resuming run 17 to step 4000
+[13:19:13] resumed: step 1004 logged, 3 trainer proc(s)
+[13:19:13] LR after resume: lr 1.8000e-04  (must still be ~3.0e-4, NOT ramping)