Fix double-subtraction of pos_ in TextLLMRunner::generate() (#18727)

[kirklandsign]

Summary:

When seq_len is set and pos_ > 0 (multi-turn conversations),
max_context_len was pre-adjusted by subtracting pos_, but
resolve_max_new_tokens then only subtracted num_prompt_tokens
instead of the full occupied position count. This caused
min(seq_len, max_context_len) to use a too-large max_context_len,
producing more tokens than allowed by seq_len.

Fix: use raw metadata value for max_context_len and pass pos_
(which includes prompt tokens after prefill) to
resolve_max_new_tokens, matching multimodal_runner's behavior.

Differential Revision: D99742232

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18727

• 📄 Preview Python docs built from this PR

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

❌ 3 New Failures, 3 Pending, 3 Unrelated Failures

As of commit 2332004 with merge base 19f7ff2 ():

NEW FAILURES - The following jobs have failed:

• Cadence Build & Test / cpu-test / test-ops / test-ops (gh)
  examples/cadence/operators/test_g3_ops.py::ATenOpTestCases::test_g3_neg_out_1
• pull / unittest-arm-backend-with-no-deps (test_pytest_ops_tosa) / linux-job (gh)
  RuntimeError: Command docker exec -t ad481c7b9b4710b6cb920511b73051c1075e61efacf9d3d61f81fd2479967f62 /exec failed with exit code 1
• pull / unittest-editable / macos / macos-job (gh)
  export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

• pull / test-models-linux (emformer_join, portable, linux.4xlarge.memory) / linux-job (gh) (matched linux rule in flaky-rules.json)
  The runner has received a shutdown signal. This can happen when the runner service is stopped, or a manually started runner is canceled.
• pull / test-models-linux (resnet18, xnnpack-quantization-delegation, linux.2xlarge) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)
• Test Vulkan Backend / test-vulkan / package-golden-artifacts (gh) (detected as infra flaky with no log or failing log classifier)

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

[meta-codesync]

@kirklandsign has exported this pull request. If you are a Meta employee, you can view the originating Diff in D99742232.

[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.

[Copilot]

Pull request overview

Fixes token budget resolution in TextLLMRunner::generate() for multi-turn conversations when seq_len is set, aligning behavior with multimodal_runner so occupied KV-cache positions are correctly accounted for.

Changes:

• Stop pre-adjusting max_context_len by pos_; use the raw metadata value instead.
• Resolve max_new_tokens using the full occupied position count (pos_ after prefill), and tighten the max-context prefill guard accordingly.
• Add a regression test covering the multi-turn + seq_len case.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
extension/llm/runner/text_llm_runner.cpp	Corrects max token resolution by using raw max_context_len and passing occupied positions (pos_) into resolve_max_new_tokens().
extension/llm/runner/test/test_text_llm_runner.cpp	Adds regression coverage to ensure seq_len limits respect prior-turn pos_ occupancy.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

GenerationConfig::resolve_max_new_tokens() takes int32_t parameters documented as num_prompt_tokens, but this call passes pos_ (an int64_t occupied-position count). This relies on implicit narrowing conversions and on a broader interpretation of the parameter than the API/docstring (and the pybinding arg name) suggests. Consider updating resolve_max_new_tokens to accept an int64_t occupied token count (or adding a new helper with clearer naming) and adjusting the documentation/bindings to avoid truncation risk and confusion.

[larryliu0820]

Review automatically exported from Phabricator review in Meta.