Upload folder using huggingface_hub

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_absent_b0.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_absent_b1.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_absent_b2.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_absent_b3.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_absent_b3.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_arm_c_b8/csv/ppl_needle_T16384_absent_b3.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_arm_c_b8/csv/ppl_needle_T16384_absent_b3.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_absent_b4.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_absent_b4.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_arm_c_b8/csv/ppl_needle_T16384_absent_b4.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_arm_c_b8/csv/ppl_needle_T16384_absent_b4.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_absent_b5.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_absent_b5.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_arm_c_b8/csv/ppl_needle_T16384_absent_b5.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_arm_c_b8/csv/ppl_needle_T16384_absent_b5.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_absent_b6.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_absent_b6.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_arm_c_b8/csv/ppl_needle_T16384_absent_b6.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_arm_c_b8/csv/ppl_needle_T16384_absent_b6.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_absent_b7.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_absent_b7.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_arm_c_b8/csv/ppl_needle_T16384_absent_b7.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_arm_c_b8/csv/ppl_needle_T16384_absent_b7.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_null_b0.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_null_b1.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_null_b2.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_null_b3.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_null_b3.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_arm_c_b8/csv/ppl_needle_T16384_null_b3.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_arm_c_b8/csv/ppl_needle_T16384_null_b3.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_null_b4.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_null_b4.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_arm_c_b8/csv/ppl_needle_T16384_null_b4.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_arm_c_b8/csv/ppl_needle_T16384_null_b4.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_null_b5.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_null_b5.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_arm_c_b8/csv/ppl_needle_T16384_null_b5.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_arm_c_b8/csv/ppl_needle_T16384_null_b5.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_null_b6.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_null_b6.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_arm_c_b8/csv/ppl_needle_T16384_null_b6.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_arm_c_b8/csv/ppl_needle_T16384_null_b6.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_null_b7.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_null_b7.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_arm_c_b8/csv/ppl_needle_T16384_null_b7.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_arm_c_b8/csv/ppl_needle_T16384_null_b7.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_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_arm_c_b8/csv/ppl_needle_T16384_p0.25_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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b0.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_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_arm_c_b8/csv/ppl_needle_T16384_p0.25_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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b1.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_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_arm_c_b8/csv/ppl_needle_T16384_p0.25_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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b2.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_b3.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_b3.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b3.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b3.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_b4.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_b4.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b4.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b4.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_b5.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_b5.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b5.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b5.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_b6.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_b6.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b6.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b6.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.25_b7.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.25_b7.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b7.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_arm_c_b8/csv/ppl_needle_T16384_p0.25_b7.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_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_arm_c_b8/csv/ppl_needle_T16384_p0.5_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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b0.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_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_arm_c_b8/csv/ppl_needle_T16384_p0.5_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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b1.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_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_arm_c_b8/csv/ppl_needle_T16384_p0.5_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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b2.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_b3.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_b3.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b3.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b3.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_b4.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_b4.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b4.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b4.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_b5.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_b5.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b5.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b5.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_b6.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_b6.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b6.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b6.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p0.5_b7.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p0.5_b7.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b7.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_arm_c_b8/csv/ppl_needle_T16384_p0.5_b7.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_p1_b0.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_p1_b1.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/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_arm_c_b8/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_arm_c_b8/csv/ppl_needle_T16384_p1_b2.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p1_b3.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p1_b3.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_arm_c_b8/csv/ppl_needle_T16384_p1_b3.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_arm_c_b8/csv/ppl_needle_T16384_p1_b3.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p1_b4.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p1_b4.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_arm_c_b8/csv/ppl_needle_T16384_p1_b4.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_arm_c_b8/csv/ppl_needle_T16384_p1_b4.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p1_b5.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p1_b5.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_arm_c_b8/csv/ppl_needle_T16384_p1_b5.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_arm_c_b8/csv/ppl_needle_T16384_p1_b5.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p1_b6.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p1_b6.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_arm_c_b8/csv/ppl_needle_T16384_p1_b6.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_arm_c_b8/csv/ppl_needle_T16384_p1_b6.csv (16384 positions, 8 windows)

logs/probe_arm_c_b8/csv/1b_run8_T16384_yarn_p1_b7.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_run16_step_400.ckpt  (base step 400, loss 1.53669)  -- 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_log1.csv (run 1)
+[rope_theta] 500000 compiled default (no sidecar -- legacy run, assuming it was trained at this base)
+Checkpointing: ON -> checkpoints_bluscriptcp/blumodelcp_run1_step_<N>.ckpt (every 50 steps, keep 6)
+[eval] shared token stream: 131073 int32 tokens from probe_arm_c_b8/needles/needle_T16384_p1_b7.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_arm_c_b8/csv/ppl_needle_T16384_p1_b7.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_arm_c_b8/csv/ppl_needle_T16384_p1_b7.csv (16384 positions, 8 windows)