jax-ml/jax · error · ValueError
gpu.launch op not found.
Error message
gpu.launch op not found.
What it means
Raised by _gpu_launch_op when no gpu.LaunchOp is found anywhere in the module being lowered. Mosaic GPU lowering expects exactly one kernel launch in the module; its absence indicates a malformed kernel.
Solutions
- Use the standard entry point (e.g. mosaic gpu kernel launch utilities) that creates the gpu.launch wrapper before lowering
- If building IR manually, wrap your func in a gpu.launch op before invoking the lowering
- Check pass ordering — ensure launch materialization runs before the lowering pass
Defensive patterns
Strategy: validation
Validate before calling
assert any(isinstance(o, gpu.LaunchOp) for f in module.body.operations for r in f.operation.regions for b in r.blocks for o in b.operations), 'missing gpu.launch'
Prevention
- Always build kernels through the standard mosaic launch utilities
- Verify pass ordering before calling lower_mosaic_module on custom IR
When it happens
Trigger: Calling the lower_mosaic_module pass on a module whose top-level function contains no gpu.launch op — e.g. hand-built IR, partially constructed modules, or a pass ordering where the launch wrapper hasn't been created yet.
Common situations: Running lower_mosaic_module directly on user IR without going through the normal kernel-build path; incorrect pass pipeline ordering after refactor/version changes.
Related errors
- A custom return op must terminate the block.
- has a memory space that is not TMEM.
- is not a memref.
- 2D+ vectors are not supported in debug_print
- 4-bit block scaled MMA only supports K-fastest operands…
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/faa0f339cadcb3b5.
Report an issue: GitHub.
Appendix: source
Thrown at jax/experimental/mosaic/gpu/dialect_lowering.py:2985
return (
# pyrefly: ignore[missing-attribute]
op.OPERATION_NAME.startswith("mosaic_gpu.")
or inference_utils.should_have_layout(op)
or inference_utils.should_have_transforms(op)
or inference_utils.should_have_tmem_layout(op)
# Does it have subblocks?
or any(bool(b) for r in op.regions for b in r)
)
def _gpu_launch_op(module: ir.Module) -> gpu.LaunchOp:
for op in module.body.operations:
for region in op.operation.regions:
for block in region.blocks:
for sub_op in block.operations:
if isinstance(sub_op, gpu.LaunchOp):
return sub_op
raise ValueError("gpu.launch op not found.")
def _lowering_context(
module: ir.Module,
launch_context: lc.LaunchContext | None,
auto_barriers: bool,
) -> LoweringContext:
"""Returns a `LoweringContext` for the given `LaunchContext`."""
# TODO(bchetioui): fix tests to not have a test-only path polluting the API.
if launch_context is None: # this case is used in some tests
return LoweringContext(None, None, None, None, None, auto_barriers, 10**9)
gpu_launch_op = _gpu_launch_op(module)
with ir.InsertionPoint.at_block_begin(gpu_launch_op.regions[0].blocks[0]):
eq = arith.CmpIPredicate.eq
i32 = ir.IntegerType.get_signless(32)
single_warp_per_block_predicate = arith.cmpi(
eq, utils.warp_idx(sync=False), utils.c(0, i32)View on GitHub (pinned to 1e1c6a8fc0)