Upload folder using huggingface_hub

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_absent_b0.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_absent_b0.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_absent_b0.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_absent_b0.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_absent_b1.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_absent_b1.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_absent_b1.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_absent_b1.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_absent_b2.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_absent_b2.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_absent_b2.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_absent_b2.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_null_b0.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_null_b0.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_null_b0.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_null_b0.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_null_b1.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_null_b1.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_null_b1.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_null_b1.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_null_b2.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_null_b2.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_null_b2.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_null_b2.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_p0.6670_b0.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_p0.6670_b0.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_p0.6670_b0.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_p0.6670_b0.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_p0.6670_b1.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_p0.6670_b1.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_p0.6670_b1.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_p0.6670_b1.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_p0.6670_b2.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_p0.6670_b2.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_p0.6670_b2.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_p0.6670_b2.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_p1_b0.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_p1_b0.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_p1_b0.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_p1_b0.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_p1_b1.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_p1_b1.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_p1_b1.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_p1_b1.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/csv/1b_run8_T16384_yarn_p1_b2.log

@@ -0,0 +1,43 @@
+=== Llama Context Parallel Training Script (bluscriptCP) ===
+[CP attn fusion] split (Stage-2 rope-once + Stage-3 Hopper causal-attn kernels)
+================ Configuration (bluscriptCP) ================
+  d_model/layers  : 1536 / 36
+  q/kv heads      : 12 / 4  (G=3, hd=128)
+  ffn_hidden      : 4096 (SwiGLU)
+  context_length  : 16384
+  vocab_size      : 50304  tying=0
+  B x T           : 1 x 16384
+  world_size      : 1  (dp=1 x cp=1)
+  CP_ATTN_MODE    : ring  (load_balance=HeadTail)
+  global_batch    : 524288  (grad_accum=32)
+  optimizer       : AdamW  [DDP]
+  precision       : fp32 everywhere
+  max_lr/min_lr   : 0.0006 / 6e-05
+  warmup/steps    : 91 / 917  (cosine, no stable phase)
+  checkpointing   : true  (ckpt_freq=50, dir=checkpoints_bluscriptcp)
+=============================================================
+Initializing model on CUDA device 0...
+[YaRN][warn] seq_len(16384) != scale(4.000)*orig_maxpos(1024.0)=4096.0 -- cache scaled for a different context than it is sized for
+[YaRN] build_rope_cache: seq_len=16384 hd=128 scale=4.000 low=7 high=25 m=1.1386
+[InitFrom] weights loaded from /root/Adhi/BluTrain/dist/Context_Parallelism/checkpoints_bluscriptcp/rank_0/blumodelcp_run15_step_400.ckpt  (base step 400, loss 1.45028)  -- optimizer state and RNG intentionally NOT restored; pair with CP_REWARMUP
+Parameters: 1060624896
+Parameters per GPU: 1060624896
+max_steps: 917
+warmup_steps: 91
+found 100 shards for split train
+found 1 shards for split val
+Saving logs to: CP_BluScreipt_Training_logs/CP_Training_log9.csv (resume run 9)
+[rope_theta] 500000 from checkpoints_bluscriptcp/blumodelcp_run9.meta (authoritative)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run9_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_pos12288_armB/needles/needle_T16384_p1_b2.bin (B forced to 1 for HF-scorer parity)
+[eval] PPL-vs-position: 8 windows (skip=0) x B=1 Tglobal=16384 Tlocal=16384 cp=1 (YARN_SCALE via env) -> probe_pos12288_armB/csv/ppl_needle_T16384_p1_b2.csv
+[eval] per-position NLL: GPU
+[eval] window 1/8 done
+[eval] window 2/8 done
+[eval] window 3/8 done
+[eval] window 4/8 done
+[eval] window 5/8 done
+[eval] window 6/8 done
+[eval] window 7/8 done
+[eval] window 8/8 done
+[eval] wrote probe_pos12288_armB/csv/ppl_needle_T16384_p1_b2.csv (16384 positions, 8 windows)

logs/probe_pos12288_armB/needles/lam_yarn_hd128_th500000_s4.txt

@@ -0,0 +1,4 @@
+n_hat 0
+s 4
+S_search 16384
+lambda 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.04 1.09 1.14 1.20 1.26 1.33 1.41 1.50 1.60 1.71 1.85 2.00 2.18 2.40 2.67 3.00 3.43 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00 4.00

logs/probe_pos12288_armB/needles/needle_T16384_absent_b0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b55db9d0e6310e4f2b2450d471978d427ace17fe6f5264da82829cae4f87b44
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_absent_b0.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_absent_b0.bin",
+  "variant": "absent",
+  "block": 0,
+  "p_frac": null,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    null,
+    null,
+    null,
+    null,
+    null,
+    null,
+    null,
+    null
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_absent_b1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4cf76184d68a874c66e7ef78ceb1160c98a24f3fe7cab21114af9b24203c166
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_absent_b1.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_absent_b1.bin",
+  "variant": "absent",
+  "block": 1,
+  "p_frac": null,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    null,
+    null,
+    null,
+    null,
+    null,
+    null,
+    null,
+    null
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_absent_b2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ed3685de1160b11e0ab73bc74290db4f62d5842eda94409368673077e4f09ce
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_absent_b2.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_absent_b2.bin",
+  "variant": "absent",
+  "block": 2,
+  "p_frac": null,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    null,
+    null,
+    null,
+    null,
+    null,
+    null,
+    null,
+    null
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_null_b0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b3de1304ff6236c02e188c9ba5bc20a005cf8ada348a273e848a4feb1fccf95
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_null_b0.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_null_b0.bin",
+  "variant": "null",
+  "block": 0,
+  "p_frac": 0.0,
+  "is_null": true,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_null_b1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4199c423878bbefd7147988f4515fdaff3855006a24f607ebda989bdaea7502
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_null_b1.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_null_b1.bin",
+  "variant": "null",
+  "block": 1,
+  "p_frac": 0.0,
+  "is_null": true,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_null_b2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:636f2fdbc8140e90482a6d686c7039339b485a2f6efe617aaa194d7676877eb6
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_null_b2.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_null_b2.bin",
+  "variant": "null",
+  "block": 2,
+  "p_frac": 0.0,
+  "is_null": true,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_p0.6670_b0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0018af943fad3a024590891ce00f5fcc56470e5262aa9e09aec955dc865615ce
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_p0.6670_b0.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_p0.6670_b0.bin",
+  "variant": "p0.6670",
+  "block": 0,
+  "p_frac": 0.667,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_p0.6670_b1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c300a3d79da4ea9641b4381ae7ad7a0740b97b78716509d71f1c1cb4c5f10b25
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_p0.6670_b1.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_p0.6670_b1.bin",
+  "variant": "p0.6670",
+  "block": 1,
+  "p_frac": 0.667,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_p0.6670_b2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ead506f1c6e31da2ea1bb5488b976f3ec23cccf8b014ed41c71a473c96816c6
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_p0.6670_b2.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_p0.6670_b2.bin",
+  "variant": "p0.6670",
+  "block": 2,
+  "p_frac": 0.667,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185,
+    8185
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_p1_b0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90ee8396a417f546557c366232048873508bf59283d9abfa1660280719c12ecd
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_p1_b0.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_p1_b0.bin",
+  "variant": "p1",
+  "block": 0,
+  "p_frac": 1.0,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_p1_b1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddf9ec99d389af7e0280a2b9708b9d16a823c1ae10fc6db0e997a9ae5fd1b16b
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_p1_b1.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_p1_b1.bin",
+  "variant": "p1",
+  "block": 1,
+  "p_frac": 1.0,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/needles/needle_T16384_p1_b2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54f627d7a04474cde21f353a5f22432adfb37b52ee4f318ff873eacd7959de47
+size 524292

logs/probe_pos12288_armB/needles/needle_T16384_p1_b2.json

@@ -0,0 +1,37 @@
+{
+  "bin": "probe_pos12288_armB/needles/needle_T16384_p1_b2.bin",
+  "variant": "p1",
+  "block": 2,
+  "p_frac": 1.0,
+  "is_null": false,
+  "trials": 8,
+  "seed": 0,
+  "vocab": 50304,
+  "reverse": false,
+  "slot_starts": [
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272,
+    12272
+  ],
+  "T": 16384,
+  "needle_len": 16,
+  "pad_tail": 4,
+  "probe_pos": 12288,
+  "probe_start": 12288,
+  "answer_start": 12296,
+  "answer_end": 12304,
+  "max_slot_start": 12272,
+  "answer_offsets": [
+    12296,
+    12304
+  ],
+  "csv_positions": [
+    12295,
+    12303
+  ]
+}

logs/probe_pos12288_armB/reach.csv

@@ -0,0 +1,3 @@
+model,rope_theta,head_dim,T,factor,cache,p_frac,nll,se,nll_null,se_null,nll_absent,nll_floor,denom_nats,score,significant,n_trials_tokens,blocks,notes
+1b_run8,500000,128,16384,4,yarn,p0.6670,3.807108,0.526386,13.789833,0.166538,13.857169,1.286227,12.570942,0.7995,1,64,3,
+1b_run8,500000,128,16384,4,yarn,p1,1.286227,0.103496,13.789833,0.166538,13.857169,1.286227,12.570942,1.0000,1,64,3,