{"record":{"id":"a38d1afc609fb38a","repo":"jax-ml/jax","slug":"num-seqs-0-must-be-less-or-equal-to-max-num-s","errorCode":null,"errorMessage":"{num_seqs[0]=} must be less or equal to {max_num_seqs=}","messagePattern":"(.+?) must be less or equal to (.+?)","errorType":"validation","errorClass":"ValueError","httpStatus":null,"severity":"error","filePath":"jax/experimental/pallas/ops/tpu/ragged_paged_attention/kernel.py","lineNumber":193,"sourceCode":"      kv_lens,\n      page_indices,\n      cu_q_lens,\n      num_seqs,\n      sm_scale=sm_scale,\n      sliding_window=sliding_window,\n      soft_cap=soft_cap,\n      mask_value=mask_value,\n      k_scale=k_scale,\n      v_scale=v_scale,\n      num_kv_pages_per_block=num_kv_pages_per_block,\n      num_queries_per_block=num_queries_per_block,\n      vmem_limit_bytes=vmem_limit_bytes,\n  )\n  max_num_batched_tokens = q.shape[0]\n  page_size = kv_pages.shape[1]\n  max_num_seqs, pages_per_seq = page_indices.shape\n  if num_seqs[0] > max_num_seqs:\n    raise ValueError(f\"{num_seqs[0]=} must be less or equal to {max_num_seqs=}\")\n  max_kv_len = jnp.max(kv_lens)\n  min_pages_per_seq = pl.cdiv(max_kv_len, page_size)\n  if pages_per_seq < min_pages_per_seq:\n    raise ValueError(\n        f\"{pages_per_seq=} must be greater or equal to\"\n        f\" {min_pages_per_seq=} given {max_kv_len=} and {page_size=}.\"\n    )\n  if cu_q_lens[num_seqs[0]] > max_num_batched_tokens:\n    raise ValueError(\n        f\"Total q tokens {cu_q_lens[num_seqs[0]]} must be less or equal to\"\n        f\" {max_num_batched_tokens=}.\"\n    )\n  for i in range(num_seqs[0]):\n    q_len = cu_q_lens[i + 1] - cu_q_lens[i]\n    kv_len = kv_lens[i]\n    if q_len > kv_len:\n      raise ValueError(\n          f\"{q_len=} must be less or equal to {kv_len=} at sequence {i}.\"","sourceCodeStart":175,"sourceCodeEnd":211,"githubUrl":"https://github.com/jax-ml/jax/blob/1e1c6a8fc06dfcd1247076ec5cae4640cea5d7bb/jax/experimental/pallas/ops/tpu/ragged_paged_attention/kernel.py#L175-L211","documentation":"In the ragged paged-attention kernel, num_seqs is a device array holding the live number of sequences, while page_indices is preallocated to max_num_seqs rows. The dynamic (compile-time-run) validation enforces num_seqs[0] <= max_num_seqs so the kernel never reads past the page table.","triggerScenarios":"Calling ragged_paged_attention (or its _test helper via dynamic_validate_inputs) with num_seqs[0] greater than page_indices.shape[0], e.g. scheduling 40 sequences into a page table sized for 32.","commonSituations":"Continuous batching schedulers that admit more sequences than the preallocated KV-cache capacity; miscomputing max_num_seqs when allocating page_indices; shrinking cache size while keeping scheduler limits unchanged.","solutions":["Cap admitted sequences: num_seqs = min(num_seqs, page_indices.shape[0])","Preallocate page_indices with max_num_seqs >= scheduler max batch size","Validate num_seqs[0] <= page_indices.shape[0] before launching"],"exampleFix":"// before\npage_indices = jnp.zeros((32, pages), jnp.int32)\nragged_paged_attention(q, k_pages, page_indices, num_seqs=jnp.array([40]), ...)\n// after\npage_indices = jnp.zeros((64, pages), jnp.int32)  # >= max batch","handlingStrategy":"validation","validationCode":"max_num_seqs = page_indices.shape[0]\nassert int(num_seqs[0]) <= max_num_seqs, (num_seqs[0], max_num_seqs)","typeGuard":null,"tryCatchPattern":null,"preventionTips":["Size page_indices rows to the scheduler's max batch, not current batch","Centralize capacity checks in the scheduler admission step"],"tags":["jax","pallas","tpu","ragged-attention","capacity-validation"],"backgroundTag":"capacity-exceeded","analyzedSha":"1e1c6a8fc06dfcd1247076ec5cae4640cea5d7bb","analyzedAt":"2026-08-27T09:53:25.647Z","schemaVersion":2},"datasetVersion":"2026-08-27T13:17:12.746Z"}