Install a per-spec fast launcher that bypasses Triton's JITFunction.run

[yushangdi]

Stacked PRs:

• ->Install a per-spec fast launcher that bypasses Triton's JITFunction.run #2749
• Add a SymInt-free tensor specialization key for exact torch.Tensor args #2759

---

Install a per-spec fast launcher that bypasses Triton's JITFunction.run

Every Helion kernel launch went through Triton's full JITFunction.run
pipeline (~9.3us): per-call device + stream proxy resolution, the
argument binder, compute_cache_key, kernel-cache lookup, the
used_global_vals walk, launch_metadata construction (even with no
profiler attached), and the kwargs-dict munging around all of it. For
a Helion kernel almost all of that is redundant: BoundKernel has
already specialized on dtype/shape/stride/device, so the only
Triton-level specialization left at launch time is pointer alignment
and binary-affecting knob state.

This ports the _FastLauncher design from upstream PR #2565 (plus the
set_config function-clone fix from PR #2635) onto main:

• helion/runtime/_fast_launcher.py -- default_launcher moves here
  unchanged (still re-exported from helion.runtime), joined by
  _FastLauncher: a multi-spec launcher primed on first call. The hot
  path computes a tiny spec key inline -- an alignment bitmask over
  the tensor args (data_ptr() & 15) plus debug/instrumentation_mode/
  stages-hook knob state -- dict-looks-up the compiled binary for
  that spec, and jumps straight into Triton's C launcher
  (CompiledKernel.run). Spec misses compile through Triton's full
  pipeline once and are cached, so call sites alternating aligned/
  unaligned tensors stay on the fast path for both.

• BoundKernel.set_config clones the PyCodeCache'd host function
  (PyCodeCache keys on source hash, so two BoundKernels can share one
  function object) and re-points its _launcher kwdefault at a
  _FastLauncher with the config's num_warps/num_stages/etc. baked in.
  Explicit _launcher= callers (the autotune trial harness) override
  the kwdefault naturally.

• TritonBackend.launcher_runtime_kwargs factors the runtime kwarg
  values out of launcher_keyword_args so codegen strings and the
  launcher closure share one source of truth.

Safety / correctness guards, each pinned by a test in
test/test_fast_launcher.py:

• Alignment is part of the spec key, so an unaligned tensor after an
  aligned prime gets its own correctly-compiled binary -- never the
  vectorized aligned binary (which would fault), and never a clone
  (which would silently drop writes to output args).
• used_global_vals snapshot per spec entry; any mutation falls back
  to JITFunction.run so Triton's own RuntimeError surfaces instead of
  silently launching a stale binary.
• torch.compile tracing routes through default_launcher so Dynamo's
  triton_kernel_wrapper_mutation HOP rules apply.
• Multi-device guard: a current-device change after priming falls
  back to Triton's per-device dispatch.
• launch_enter/exit hooks are re-read per call (a profiler attached
  after priming still fires; launch_metadata is built only when a
  hook will consume it); pre_run_hooks fire inline; flipping
  knobs.runtime.debug lands on a new spec entry and recompiles.
• Any priming/compile failure, and HELION_SKIP_FAST_LAUNCHER=1, fall
  back to default_launcher permanently.

Verified: test_fast_launcher + test_misc (51 passed), and full runs of
test_torch_compile (244), test_examples (96), test_autotuner (122),
test_indexing, test_loops, test_grid, test_ref_eager, test_specialize,
test_config_api, test_cache earlier on this branch.

Benchmark (B200, end-to-end wall time per call, add-style kernel,
N=4096 fp32, HELION_AUTOTUNE_EFFORT=none, steady state, 20k iters):

  n_args | baseline | prev commit | this commit | total
  -------+----------+-------------+-------------+------
       2 | 24.14 us |    17.25 us |    13.63 us |  -44%
       8 | 33.05 us |    24.50 us |    18.99 us |  -43%
      16 | 43.56 us |    32.16 us |    25.73 us |  -41%

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com