
phase 1 backward microbench (B=32, T=1024, D=512, BT=32768, BLOCK_SIZE=8)
zero_each_run=False

case                                        triton full  triton main   triton red    torch bwd
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_kernel,
with key as (1, 512, 8, 1, 'torch.bfloat16', 'torch.bfloat16', 'torch.bfloat16', 'torch.float32'),
finished after 2.47s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None;
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_backward_kernel,
with key as (1, 512, 8, 1, 'torch.bfloat16', 'torch.bfloat16', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32'),
finished after 21.61s,
best config selected: num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None;
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 16, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 16, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 32, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 16, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 16, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 32, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 16, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 16, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 32, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Triton autotuning for function phase_1_reduce_grad_pseudo_queries_kernel,
with key as (32768, 512, 8, 'torch.float32', 'torch.float32'),
finished after 2.88s,
best config selected: BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None;
block_start=0  src=1  q=8  grad_lse=True          1.303        0.917        0.385        2.695
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_kernel,
with key as (2, 512, 8, 2, 'torch.bfloat16', 'torch.bfloat16', 'torch.bfloat16', 'torch.float32'),
finished after 2.51s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None;
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_backward_kernel,
with key as (2, 512, 8, 2, 'torch.bfloat16', 'torch.bfloat16', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32'),
finished after 33.53s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None;
block_start=8  src=2  q=8  grad_lse=True          1.442        1.054        0.386        4.453
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_kernel,
with key as (3, 512, 8, 4, 'torch.bfloat16', 'torch.bfloat16', 'torch.bfloat16', 'torch.float32'),
finished after 2.56s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None;
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_backward_kernel,
with key as (3, 512, 8, 4, 'torch.bfloat16', 'torch.bfloat16', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32'),
finished after 43.88s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None;
block_start=16 src=3  q=8  grad_lse=True          1.621        1.229        0.385        5.647
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_kernel,
with key as (4, 512, 8, 4, 'torch.bfloat16', 'torch.bfloat16', 'torch.bfloat16', 'torch.float32'),
finished after 2.56s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None;
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_backward_kernel,
with key as (4, 512, 8, 4, 'torch.bfloat16', 'torch.bfloat16', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32', 'torch.float32'),
finished after 52.33s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None;
block_start=24 src=4  q=8  grad_lse=True          1.814        1.418        0.386        6.778
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_kernel,
with key as (5, 512, 1, 8, 'torch.bfloat16', 'torch.bfloat16', 'torch.bfloat16', 'torch.float32'),
finished after 2.43s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None;
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 1, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 2, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 4, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 8, num_ctas: 1, num_stages: 4, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 2, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 3, maxnreg: None
Autotuning kernel phase_1_batched_interblock_attention_backward_kernel with config num_warps: 16, num_ctas: 1, num_stages: 4, maxnreg: None
Triton autotuning for function phase_1_batched_interblock_attention_backward_kernel,
with key as (5, 512, 1, 8, 'torch.bfloat16', 'torch.bfloat16', 'torch.float32', 'torch.bfloat16', 'torch.float32', 'torch.float32', 'torch.float32'),
finished after 14.35s,
best config selected: num_warps: 1, num_ctas: 1, num_stages: 3, maxnreg: None;
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 16, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 16, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 64, BLOCK_HIDDEN: 32, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 16, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 16, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 128, BLOCK_HIDDEN: 32, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 16, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 16, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None
Autotuning kernel phase_1_reduce_grad_pseudo_queries_kernel with config BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 32, num_warps: 8, num_ctas: 1, num_stages: 1, maxnreg: None
Triton autotuning for function phase_1_reduce_grad_pseudo_queries_kernel,
with key as (32768, 512, 1, 'torch.float32', 'torch.float32'),
finished after 2.67s,
best config selected: BLOCK_BATCH_SEQ: 256, BLOCK_HIDDEN: 32, num_warps: 4, num_ctas: 1, num_stages: 1, maxnreg: None;
final src=5  q=1  grad_lse=False                  0.749        0.695        0.056        2.276


phase 1 backward sweep only
Includes final phase-1 backward plus all per-block phase-1 backward calls.
triton zero_each_run=True
triton phase1 backward sweep: 7.170 ms
torch  phase1 backward sweep: 21.863 ms

