From b704557813bd58b32f481f2934a81e6702d4f0b3 Mon Sep 17 00:00:00 2001
From: Min Guo <minguo@meta.com>
Date: Thu, 22 Oct 2026 14:53:19 -0700
Subject: [PATCH] Fix `llama_qnn` Buck target import failures

Summary:
`//executorch/examples/qualcomm/oss_scripts/llama:llama_qnn` currently fails at import time for every model, before any argument is parsed, with `ModuleNotFoundError: No module named 'executorch.examples.models.gemma4'` and then `... granite_speech`. The Buck targets drifted from the sources synced from upstream: the Gemma4 enablement (#20979) added an unconditional `executorch.examples.models.gemma4` import with no owning target for `gemma4/__init__.py`, Granite-Speech added `model/audio_encoder.py` plus a `granite_speech` import, and QAD (#21036) added `quantize/qat.py`, `dataset/{constants,targets}.py` and the `train/` package, none of which are in any target.

This adds a `gemma4` library owning `gemma4/__init__.py`, a new `train` library (including its `train/config/*.yaml` resources), the missing sources in `static_llama`, `dataset` and `quantize`, and the `gemma4` and `granite_speech` deps on `llama_lib`. The `fbcode/` and `xplat/` copies are kept byte-identical.

Known limitation, not addressed here: under Buck `QNN_SDK_ROOT` is a placeholder, so `is_qnn_sdk_version_less_than("2.35")` cannot detect the SDK version and masked softmax is disabled during export even though the pinned SDK (2.50) supports it.

Authored with Claude Code.

Differential Revision: D127006045
---
 examples/models/gemma4/BUCK                   |  9 ++++++
 .../qualcomm/oss_scripts/llama/targets.bzl    | 30 +++++++++++++++++++
 2 files changed, 39 insertions(+)

diff --git a/examples/models/gemma4/BUCK b/examples/models/gemma4/BUCK
index 19f0ff90c93..9790b60acf1 100644
--- a/examples/models/gemma4/BUCK
+++ b/examples/models/gemma4/BUCK
@@ -8,6 +8,15 @@ non_fbcode_target(_kind = define_common_targets,)
 
 fbcode_target(_kind = define_common_targets,)
 
+fbcode_target(_kind = runtime.python_library,
+    name = "gemma4",
+    srcs = ["__init__.py"],
+    _is_external_target = True,
+    base_module = "executorch.examples.models.gemma4",
+    deps = [":text_decoder"],
+    visibility = ["PUBLIC"],
+)
+
 # Text decoder module
 fbcode_target(_kind = runtime.python_library,
     name = "text_decoder",
diff --git a/examples/qualcomm/oss_scripts/llama/targets.bzl b/examples/qualcomm/oss_scripts/llama/targets.bzl
index d52bdcfbc8a..54036076988 100644
--- a/examples/qualcomm/oss_scripts/llama/targets.bzl
+++ b/examples/qualcomm/oss_scripts/llama/targets.bzl
@@ -74,6 +74,7 @@ def define_common_targets(is_fbcode = False):
             srcs = [
                 "model/__init__.py",
                 "model/apply_rope.py",
+                "model/audio_encoder.py",
                 "model/embedding.py",
                 "model/feed_forward.py",
                 "model/layernorm.py",
@@ -82,6 +83,7 @@ def define_common_targets(is_fbcode = False):
             ],
             deps = [
                 "//caffe2:torch",
+                "//executorch/examples/qualcomm:utils",
                 "//executorch/examples/models/llama:transformer_modules",
                 "fbsource//third-party/pypi/transformers:transformers",
             ],
@@ -174,10 +176,12 @@ def define_common_targets(is_fbcode = False):
                 "dataset/builders.py",
                 "dataset/collators.py",
                 "dataset/config.py",
+                "dataset/constants.py",
                 "dataset/datasets.py",
                 "dataset/loaders.py",
                 "dataset/preprocessors.py",
                 "dataset/schema.py",
+                "dataset/targets.py",
             ],
             deps = [
                 ":decoder_constants",
@@ -196,17 +200,41 @@ def define_common_targets(is_fbcode = False):
             srcs = [
                 "quantize/__init__.py",
                 "quantize/ptq.py",
+                "quantize/qat.py",
                 "quantize/strategy.py",
             ],
             deps = [
                 ":decoder_constants",
                 ":inference",
+                ":train",
                 ":utils",
                 "//caffe2:torch",
                 "//executorch/backends/qualcomm/_passes:passes",
             ],
         )
 
+        runtime.python_library(
+            name = "train",
+            srcs = [
+                "train/__init__.py",
+                "train/config/__init__.py",
+                "train/config/config.py",
+                "train/loss.py",
+                "train/trainer.py",
+                "train/utils.py",
+            ],
+            resources = [
+                "train/config/lr_config.yaml",
+                "train/config/qad.yaml",
+            ],
+            deps = [
+                ":dataset",
+                "//caffe2:torch",
+                "fbsource//third-party/pypi/pyyaml:pyyaml",
+                "fbsource//third-party/pypi/tqdm:tqdm",
+            ],
+        )
+
         runtime.python_library(
             name = "mix_precision_analyzer",
             srcs = [
@@ -298,8 +326,10 @@ def define_common_targets(is_fbcode = False):
                 "//executorch/examples/models/codegen:codegen",
                 "//executorch/examples/models/gemma:gemma",
                 "//executorch/examples/models/gemma2:gemma2",
+                "//executorch/examples/models/gemma4:gemma4",
                 "//executorch/examples/models/glm:glm",
                 "//executorch/examples/models/granite:granite",
+                "//executorch/examples/models/granite_speech:granite_speech",
                 "//executorch/examples/models/internvl3:internvl3",
                 "//executorch/examples/models/smolvlm:smolvlm",
                 "//executorch/examples/models/llama:hf_download",