[JIT] Actionable diagnostic for dynamic layout field overflow

[mustafayildirim]

What

Launching a kernel whose dynamic layout field exceeds its signed packed-ABI width failed with a raw struct.error: 'i' format requires -2147483648 <= number <= 2147483647 out of _LayoutPlan's pack_into (#1176). Per the maintainer analysis in the issue, the offending field is an individual dynamic dimension (e.g. shape[0] after view(-1) of a >2^31-element tensor), not the total element count.

This PR adds a launch-time diagnostic (no ABI change): every generated fill validates its dynamic fields against the plan's packed widths (shape: signed i32; stride: i32 iff use_32bit_stride, else i64) before pack_into, and raises a ValueError naming the JIT argument, field kind, dim index, value, and limit:

ValueError: argument 'OUT': dynamic shape[0] = 2500000000 exceeds the signed
int32 ABI field (max 2147483647). The dynamic-layout buffer packs this field
at a fixed width; keep the oversized dimension static
(mark_shape_dynamic/memref type), reshape/split the tensor, or file an issue
for wider dynamic dims.

The parameter name is stamped onto the plan by _build_call_state and propagated through struct fields as Struct.field, so the diagnostic works on the fast dispatch path. Validation is one compare per dynamic field per argument — only on the launch path, never in the pack fast path when values fit.

Wider dynamic dims (i64 shapes, per the second option in the issue analysis) are deliberately out of scope: they require coordinated Python-ABI + compiler-type changes and device correctness tests on large tensors.

Testing

All on MI325X (gfx942), in-pod build against LLVM e2a39f504fe:

Check	Result
Issue's launcher (2.5B-elem view(-1)) on upstream main	FAIL — raw struct.error: 'i' format requires ...
Same launcher with this patch	PASS — raises ValueError naming arg, shape[0], value, limit
New tests/unit/test_layout_field_overflow.py (4 L0 fill tests + l2_device launch repro)	5/5 PASS
Full tests/unit/ + tests/language/ pytest (gfx942)	2504 passed, 10 skipped, 0 failed

Fixes #1176

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved moderate findings affect signed-range validation, diagnostic propagation, and test safety.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 2 · 1

Open (3)

• Rejects minimum representable signed stride · New
• Propagate parameter paths through recursive ABI plans · New
• Clarify dynamic dimension remediation · New

What changed in this PR

Adds launch-time validation for dynamic layout fields before ABI packing, replacing raw struct.error failures with actionable diagnostics.

Changes:

• Validates signed shape and stride widths.
• Propagates JIT argument and struct field names.
• Adds unit and GPU regression tests.

File	Summary and review findings
tests/​unit/​test_layout_field_overflow.py	Adds overflow and launch tests. Moderate (1 vote): add CUDA availability guarding. Moderate (1 vote): mark the 10 GB repro large_shape. Nit (1 vote): pair l2_device with rocm_lower.
python/​flydsl/​expr/​struct.py	Propagates struct field labels. Moderate (2 votes): propagate top-level parameter/path information through recursive plans to avoid ambiguous diagnostics.
python/​flydsl/​compiler/​jit_function.py	Stamps JIT parameter names onto plans. Nit (1 vote): add fast-dispatch coverage asserting the actual parameter name.
python/​flydsl/​compiler/​jit_argument.py	Implements ABI-width validation and diagnostics. Moderate (2 votes): accept the full signed lower bound. Nit (3 votes): correct contradictory remediation wording. Nit (1 vote): add DLPack-backed overflow coverage.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[mustafayildirim]

Copilot review findings addressed in 30f2e0a (+ test fixes in ab52203/34a13c19):

1. Signed lower bound — overflow_report now checks the full signed range ([-2**31, 2**31-1] for i32, [-2**63, 2**63-1] for i64); a stride of -2**31 packs cleanly and is no longer rejected. Covered by test_signed_min_values_are_accepted.

2. Struct field paths — diagnostics for struct-typed parameters now compose the full field path (argument 'payload.inner.t') instead of the ambiguous 'Struct.field' (one label shared by every field of that struct type, and un-overridable by the dispatch builder since struct instances are frozen). The struct ABI recursion builds a path_suffix (nested structs included) and _build_call_state stamps the top-level parameter name via the new _stamp_plan_param_names walk. Covered by test_nested_struct_path_in_diagnostic.

3. Remediation wording — the hint no longer suggests mark_shape_dynamic (which makes the dim more dynamic, the opposite of the workaround); it now says to use a static-layout memref / leave the dim unmarked, or reshape/split the tensor.

Verified on gfx942 (MI325X): 7/7 layout-overflow tests, full tests/unit tests/language suite 2506 passed / 0 failed with a clean cache dir.