fix(runtime): allow zero-size view at any offset (Tensor::view bounds-check)

[csy0225]

Summary

Tensor::view() asserts view_offsets[i] + view_shapes[i] <= shapes[i] for every dim — but the orchestration codegen emits guards of the form (off >= shape ? 0 : min(...)) to clamp the shape dim to 0 when a dynamic offset runs past the parent, leaving the offset as the raw (possibly out-of-range) value and expecting the runtime to honour the empty-view contract. With the current assertion, that codegen pattern fails on board.

This PR loosens view() to detect any zero-shape dim up front and skip both the per-dim offset assertion and the start_offset advancement (no bytes captured, so offset bounds are irrelevant) and the trailing assert_in_buffer_bounds(). All previously valid views still pass through the same code path; only previously-asserting-but-now-OK views (zero shape + out-of-range offset) now correctly return an empty Tensor.

Repro signature on board (a2a3)

The current assertion surfaces as a generic AICPU exception errorCode=0x2a → host-side aclrtSynchronizeStreamWithTimeout (AICPU) failed: 507018. With ASCEND_GLOBAL_LOG_LEVEL=1 and ASCEND_PROCESS_LOG_PATH=<dir>, the device-side plog shows:

[ERROR] AICPU: Assertion failed:
        view_offsets[i] + view_shapes[i] <= shapes[i]
[ERROR] AICPU: Location:
        .../runtime/tensormap_and_ringbuffer/runtime/tensor.h:356

The trigger is the codegen lowering of dynamic offsets like static_cast<uint32_t>(pos) where pos = ctx_len - 1. When ctx_len is 0 (the very first prefill step, or a dummy-weight bring-up where seq_lens is not pre-populated), pos = -1 and the cast underflows to 0xFFFFFFFF. The codegen guard correctly sets cos_row_shapes[0] = 0 for this offset, but the OFFSET passed to view() is the raw 0xFFFFFFFF, which fails the assertion.

A typical generated chip-side site (paraphrased from aicpu_orchestration_entry):

int32_t ctx_len = get_tensor_data<int32_t>(ext_seq_lens, 1, indices_ctx_len);
int64_t pos = (static_cast<int64_t>(ctx_len) - 1);
uint32_t cos_row_offsets[2] = {static_cast<uint32_t>(pos), 0};
uint32_t cos_row_shapes[2] = {
    (cos_row_offsets[0] >= ext_rope_cos.shapes[0]
        ? 0u
        : std::min<uint32_t>(1, ext_rope_cos.shapes[0] - cos_row_offsets[0])),
    ...
};
Tensor cos_row = ext_rope_cos.view(cos_row_shapes, cos_row_offsets);  // <-- asserts

Why fix it in the runtime, not the codegen

The codegen's intent is unambiguous: (off >= shape ? 0 : ...) is exactly "if offset is out of range, return an empty view". Hardening every codegen call site to also clamp the offset would touch many emit paths and is prone to regression; honoring the empty-view contract once in the runtime is a strict improvement and matches what the codegen already documents in its emitted comments.

Diff shape

• Two byte-identical sibling headers (a2a3 and a5) updated symmetrically.
• ABI unchanged: sizeof(Tensor) is untouched and the static_asserts in comm_context.h continue to hold.
• No call-site changes anywhere — purely a one-method semantic widening.

Test plan

• Existing tests/ut/cpp/{a2a3,a5}/test_tensormap.cpp view-based overlap tests continue to pass (no preexisting test exercises the zero-shape-view + out-of-range-offset path).
• Suggest adding a regression test: base.view([0, ...], [0xFFFFFFFF, 0]) returns an empty Tensor with numel()==0 and does not abort.
• Validated locally on Ascend a2a3 single-rank decode: with this fix the assertion never fires; the chip-side run progresses past the rope-view sites it previously aborted at.

References

The originating model-side investigation is documented in this companion issue / mirror in the consumer project: pypto-side root-cause walkthrough exists at docs/upstream-issues/pypto-1702-followup.md (separate codepath from PR hw-native-sys/pypto#1718).

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 03b45a46-eb7f-4a61-aafd-1f3dac45dc2d

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

This PR applies identical empty-view handling logic to the Tensor::view(...) method in two parallel tensor runtime implementations. When a view has any zero-size dimension, bounds validation checks are skipped; non-empty views retain their original offset and buffer bounds assertions.

Changes

Empty view bounds checking in Tensor::view

Layer / File(s)	Summary
Empty view bounds checking skip src/a2a3/runtime/tensormap_and_ringbuffer/runtime/tensor.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/tensor.h	Tensor::view detects empty views (any dimension is zero) and skips per-dimension offset bounds debug assertions, start_offset updates, and final assert_in_buffer_bounds() calls; non-empty views preserve original validation.

🎯 2 (Simple) | ⏱️ ~8 minutes

A view that's truly empty needs no bounds—
Zero dimensions, no offset confounds!
The tensor now skips checks with grace,
For views that just vanish without trace. 🐰✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: allowing zero-size views at any offset in the Tensor::view() bounds-check logic.
Description check	✅ Passed	The description is comprehensive and directly related to the changeset, explaining the problem, solution, rationale, and validation approach for the Tensor::view() bounds-check modification.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}