[executorch][muse] Add export_solo_batching.py

[Gasoonjia]

Stack from ghstack (oldest at bottom):

• [executorch][muse] Move off-graph KV to the batching path and test batching in OSS CI #23375
• [executorch][muse] Add run_solo_batching #23374
• -> [executorch][muse] Add export_solo_batching.py #23373
• [executorch][cuda] Let CudaExecutor honor a program's prefill lower bound #23372
• [executorch][llm] Share the batching executors' session logic under extension/llm/batching/util #23371
• [executorch][cuda] Add CudaExecutor #23370
• [executorch][cuda] Add step_plan for the CUDA batching executor #23369
• [executorch][llm] Add Buck targets for extension/llm/batching #23368
• [executorch][cuda] Add CudaCellCache #23367
• [executorch][cuda] Lower off-graph KV in a cell layout #23366
• [executorch][cuda] Extract CudaKVPool from CudaSequenceKVCache #23365

Exports Muse Glimmer text generation as the artifact CudaExecutor drives. It
lives beside the single-sequence export_solo.py, which it leaves untouched.

Two methods over one set of weights, both (tokens[1, T], input_pos[T], logits_to_keep[K]) -> logits[K, vocab] float32:

• decode: T = K = 1, static; the runtime captures it into a CUDA graph.
• prefill: T and K in [5, W], dynamic. MG's quantized linears switch kernels
  at M <= 4, so a dynamic width only traces from 5, as in export_solo;
  get_min_prefill_chunk = 5 tells CudaExecutor to run narrower slices as
  decodes.

The whole step runs through the decoder so every token's K/V reaches the
cache; only the rows the batch samples pass through the LM head.

The delegate reads the step's token count from input_pos[T], the cache
ops' position, as the backend's CheckOffGraphKVStepWidthPass requires. The
spec names a delegate input, and the delegate's inputs need not follow the
method's order: decode's delegate takes (tokens, input_pos, logits_to_keep) (spec 1:0), while prefill reads logits_to_keep's
symbolic size first and takes (tokens, logits_to_keep, input_pos) (spec
2:0). Sampling happens on the host, per session, so logits are not kept on
the device.

Reuses export_solo's loaders and constant methods and the off-graph source
transform, whose manifest gains layout: cell and max_cells. Adds the
metadata CudaExecutor reads: get_max_context_len,
get_logits_to_keep_mode = selected, get_offgraph_kv_max_cells, and the
cache geometry.

CLI: --gguf | --prequantized | --checkpoint-dir, --max-seq-len (the
per-session context), --max-step-tokens (W, default 512), --max-cells
(the shared pool, default 65536; memory grows with use).

The new test also joins the OSS unittest-cuda job, where MG's CUDA tests did
not run before.

Differential Revision: D123051484

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23373

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 6 New Failures, 2 Unclassified Failures

As of commit d91f120 with merge base c16dd0a ():

NEW FAILURES - The following jobs have failed:

• pull / unittest / linux / linux-job (gh)
  backends/apple/coreml/test/test_torch_ops.py::TestTorchOps::test_dequantize_affine_c8w_embedding_c8w_linear_ios16
• pull / unittest / macos / macos-job (gh)
  backends/apple/coreml/test/test_coreml_recipes.py::TestCoreMLRecipes::test_int8_weight_only_per_group
• pull / unittest / windows / windows-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows
• pull / unittest-editable / linux / linux-job (gh)
  backends/apple/coreml/test/test_torch_ops.py::TestTorchOps::test_dequantize_affine_c8w_embedding_c8w_linear_ios16
• pull / unittest-editable / macos / macos-job (gh)
  backends/apple/coreml/test/test_coreml_recipes.py::TestCoreMLRecipes::test_int8_weight_only_per_group
• pull / unittest-editable / windows / windows-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows

UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:

• Test CUDA Builds / unittest-cuda / linux-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  backends/cuda/tests/test_offgraph_kv.py::OffGraphKVCellCompileTest::test_decode_and_prefill_from_two_tokens_lower
• Test CUDA Builds / unittest-cuda-runtime / linux-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  [ FAILED ] CudaKVCacheTest.CacheOutlivesTheStreamItLastSteppedOn

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.