Arm backend: add VGF PT2E linear quantization modes for LLM export

[xingguo01]

• add vgf_16a8w/8a8w PT2E quantization modes
• add backend.vgf.quantize_scope for full vs linear VGF quantization
• wire the VGF config through the LLM export and quantizer selection path
• add coverage in export_llama_lib tests for the new VGF PT2E modes

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19029

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Run pull request jobs on OSDC runners in shadow mode

❌ 3 New Failures, 2 Cancelled Jobs, 10 Unrelated Failures

As of commit 7d65ac8 with merge base d6f1625 ():

NEW FAILURES - The following jobs have failed:

• pull / test-sqnr-static-llm-qnn-linux (smollm2_135m) / linux-job (gh)
  test_static_llm_model
• pull / unittest-editable / linux / linux-job (gh)
  examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
• Test CoreML Backend / test-coreml / test-backend-macos (coreml, models) / macos-job (gh)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• trunk / test-huggingface-transformers-xnnpack (qwen3-1.7b|xnnpack|--quantize) / linux-job (gh)
  ##[error]The operation was canceled.
• trunk / test-qnn-model (fp32, mb) / linux-job (gh)
  ##[error]The operation was canceled.

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / unittest / macos / macos-job (gh) (similar failure)
  ##[error]The operation was canceled.

BROKEN TRUNK - The following jobs failed but was present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-models-linux (mobilebert, portable, linux.2xlarge) / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / test-models-linux (mobilebert, xnnpack-quantization-delegation, linux.2xlarge) / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / test-voxtral-realtime-xnnpack-linux / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / unittest-editable / macos / macos-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• trunk / test-models-linux-aarch64 (mobilebert, portable, linux.arm64.2xlarge) / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• trunk / test-models-linux-aarch64 (mobilebert, xnnpack-quantization-delegation, linux.arm64.2xlarge) / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• trunk / test-models-windows (mobilebert, portable) / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• trunk / test-models-windows (mobilebert, xnnpack-q8) / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• trunk / unittest-release / macos / macos-job (gh) (trunk failure)
  ##[error]The operation was canceled.

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

[Copilot]

Pull request overview

Adds Arm VGF backend PT2E quantization support for LLM export, including a new 16a8w mode gated on INT16 TOSA extension support and a configurable quantization scope (full-model vs Linear-only), plus test coverage for the new behavior.

Changes:

• Add vgf_16a8w PT2E quantization mode and enforce INT16 compile spec extension when selected.
• Introduce backend.vgf.quantize_scope (full vs linear) and apply it when constructing the VGF quantizer.
• Wire new VGF settings through llama export CLI/config and add unit tests for scope + INT16 gating.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
extension/llm/export/quantizer_lib.py	Extends VGF quantizer selection for vgf_16a8w and adds scope-based application (global vs Linear-only).
extension/llm/export/config/llm_config.py	Adds vgf_16a8w enum value and introduces VgfQuantizeScope + config wiring from CLI args.
examples/models/llama/export_llama_lib.py	Exposes VGF PT2E modes and VGF scope/compile-spec CLI flags; passes scope into VGF quantizer creation.
examples/models/llama/tests/test_export_llama_lib.py	Adds coverage for VGF linear-only scope and INT16 compile spec enforcement for vgf_16a8w.

---

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

[Copilot]

The --pt2e_quantize argparse option is defined with a fixed set of choices, so it only accepts a single value, but the help text says it supports "Comma separated options" (and even mentions embedding, which is not a valid choice). This is user-facing and likely to confuse; either update the help text to reflect single-choice behavior, or switch the argument parsing to accept a comma-separated list (and adjust LlmConfig/Pt2eQuantize parsing accordingly).

Suggested change

	help="Use PT2E quantization. Comma separated options. e.g. xnnpack_dynamic (for per channel 8 bit weight), xnnpack_dynamic_qc4 (for per channel 4 bit weight), embedding.",
	help="Use a single PT2E quantization mode, e.g. xnnpack_dynamic (per-channel 8-bit weight) or xnnpack_dynamic_qc4 (per-channel 4-bit weight).",

[xingguo01]

@digantdesai @SS-JIA highly appreciated if you can have a look

[zingo]

Mostly touch Arm backend stuff, and no need for Buck2 fixes

[digantdesai]

nit ```suggestion
parser.add_argument("--arm-vgf", action="store_true")

rationale is not everyone knows what vgf is

[digantdesai]

@claude review this PR

[claude]

Claude finished @digantdesai's task in 1m 59s —— View job

---

Review of PR #19029: Arm backend VGF PT2E linear quantization modes for LLM export

• Read PR diff and changed files
• Analyze code for correctness, style, and potential issues
• Post review feedback

---

Overall: Clean, well-structured PR. The new quantization modes are properly gated, the scope abstraction is sensible, and the tests cover the key paths (full vs linear scope, INT16 enforcement). A few observations below:

---

1. Reviewer feedback: --vgf → --arm-vgf (digantdesai's comment)

I agree with @digantdesai's suggestion at export_llama_lib.py:461. "VGF" is Arm-specific terminology and not widely recognizable. Prefixing with --arm-vgf (and correspondingly --arm-vgf-compile-spec, --arm-vgf-quantize-scope) would be more discoverable for users scanning --help. The other Arm backends (TOSA, Ethos-U) don't have this prefix either, but those are more established names — vgf is new enough that the disambiguation helps.

2. from_args silently accepts vgf_compile_spec even when --vgf is not set

In llm_config.py:828-829:

if hasattr(args, "vgf_compile_spec"):
    llm_config.backend.vgf.compile_spec = args.vgf_compile_spec

Because the argparse default is "TOSA-1.0+INT", this unconditionally overwrites the config's compile spec whether or not --vgf is passed. It's harmless today (both defaults match), but fragile if the defaults diverge. Consider gating on args.vgf or at least args.vgf_compile_spec != parser.get_default(...). Low priority since it's consistent with how other backends handle this pattern in this file.

3. Copilot's note on --pt2e_quantize help text is valid but pre-existing

The help text says "Comma separated options" but the argument uses choices= which only accepts a single value. This is a pre-existing issue unrelated to this PR — mentioning for awareness but not blocking.

4. Test coverage is good

The three new tests cover:

• vgf_8a8w + linear scope → verifies module_type_config is set
• vgf_16a8w without INT16 → verifies ValueError raised
• vgf_16a8w with INT16 → verifies success

The existing test for vgf_8a8w + full scope is also correctly extended with assertions on global_config and empty module_type_config. This gives solid boundary coverage.

5. Minor: error message could be more actionable

At quantizer_lib.py:388-389:

"vgf_16a8w requires a VGF compile spec with INT16 support, "
"for example TOSA-1.0+INT+int16."

This is clear and actionable — good.

6. Code quality

• The VgfQuantizeScope enum is a clean str, Enum which serializes nicely via .value
• The scope dispatch in get_vgf_quantizer handles invalid values with a clear error
• Import of get_symmetric_a16w8_quantization_config is properly deferred inside the function (matching existing pattern)

---

Summary: This is a solid, focused change. The only actionable feedback is the --arm-vgf naming suggestion from digantdesai. The code is correct and well-tested. LGTM once the naming nit is addressed.

---