Merge branch 'main' into remove-repeated-recompiles-in-constant-fusion

Author: YufengShi-dudu

.ci/docker/build.sh

@@ -92,18 +92,6 @@ esac
 TORCH_VERSION=$(cat ci_commit_pins/pytorch.txt)
 BUILD_DOCS=1
 
-# Pull channel + spec/url helpers out of torch_pin.py so install_pytorch.sh
-# (which runs inside the docker build, where torch_pin.py isn't available)
-# can decide between wheel install (test/release) and source build (nightly).
-# Self-hosted runners often have python3 but not the unversioned python alias.
-PYTHON_BIN=$(command -v python3 || command -v python)
-TORCH_PIN_HELPERS=$(cd ../.. && "$PYTHON_BIN" -c "from torch_pin import CHANNEL, torch_spec, torchaudio_spec, torchvision_spec, torch_index_url_base; print(CHANNEL); print(torch_spec()); print(torchaudio_spec()); print(torchvision_spec()); print(torch_index_url_base())")
-TORCH_CHANNEL=$(echo "${TORCH_PIN_HELPERS}" | sed -n '1p')
-TORCH_SPEC=$(echo "${TORCH_PIN_HELPERS}" | sed -n '2p')
-TORCHAUDIO_SPEC=$(echo "${TORCH_PIN_HELPERS}" | sed -n '3p')
-TORCHVISION_SPEC=$(echo "${TORCH_PIN_HELPERS}" | sed -n '4p')
-TORCH_INDEX_URL=$(echo "${TORCH_PIN_HELPERS}" | sed -n '5p')
-
 # Copy requirements-lintrunner.txt from root to here
 cp ../../requirements-lintrunner.txt ./
 
@@ -116,11 +104,6 @@ docker build \
   --build-arg "PYTHON_VERSION=${PYTHON_VERSION}" \
   --build-arg "MINICONDA_VERSION=${MINICONDA_VERSION}" \
   --build-arg "TORCH_VERSION=${TORCH_VERSION}" \
-  --build-arg "TORCH_CHANNEL=${TORCH_CHANNEL}" \
-  --build-arg "TORCH_SPEC=${TORCH_SPEC}" \
-  --build-arg "TORCHAUDIO_SPEC=${TORCHAUDIO_SPEC}" \
-  --build-arg "TORCHVISION_SPEC=${TORCHVISION_SPEC}" \
-  --build-arg "TORCH_INDEX_URL=${TORCH_INDEX_URL}" \
   --build-arg "BUCK2_VERSION=${BUCK2_VERSION}" \
   --build-arg "LINTRUNNER=${LINTRUNNER:-}" \
   --build-arg "BUILD_DOCS=${BUILD_DOCS}" \

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-release/2.11
+release/2.11

.ci/docker/common/install_pytorch.sh

@@ -17,24 +17,6 @@ install_domains() {
 }
 
 install_pytorch_and_domains() {
-  if [ "${TORCH_CHANNEL}" != "nightly" ]; then
-    # Test/release: install the published wheels directly. The specs and URL
-    # are passed in as docker build args (computed from torch_pin.py by
-    # .ci/docker/build.sh). RC wheels at /whl/test/ get re-uploaded under the
-    # same version, so use --no-cache-dir there to avoid stale cache hits.
-    local cache_flag=""
-    if [ "${TORCH_CHANNEL}" = "test" ]; then
-      cache_flag="--no-cache-dir"
-    fi
-    pip_install --force-reinstall ${cache_flag} \
-      "${TORCH_SPEC}" "${TORCHVISION_SPEC}" "${TORCHAUDIO_SPEC}" \
-      --index-url "${TORCH_INDEX_URL}/cpu"
-    return
-  fi
-
-  # Nightly: build pytorch from source against the pinned SHA in pytorch.txt
-  # so we catch upstream regressions, then install audio/vision from the
-  # commits that pytorch itself pins.
   git clone https://github.com/pytorch/pytorch.git
 
   # Fetch the target commit
@@ -45,19 +27,14 @@ install_pytorch_and_domains() {
   chown -R ci-user .
 
   export _GLIBCXX_USE_CXX11_ABI=1
-  # PyTorch's FindARM.cmake hard-fails when the SVE+BF16 compile probe
-  # doesn't pass — gcc-11 in this image is too old to accept the combined
-  # NEON/SVE/bfloat16 intrinsics the probe exercises. Executorch's aarch64
-  # runtime targets (phones, embedded) don't use SVE, so bypass the check.
-  export BUILD_IGNORE_SVE_UNAVAILABLE=1
   # Then build and install PyTorch
   conda_run python setup.py bdist_wheel
   pip_install "$(echo dist/*.whl)"
 
-  # Defer to PyTorch's own pinned audio/vision commits.
-  TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
+  # Grab the pinned audio and vision commits from PyTorch
+  TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt)
+  TORCHVISION_VERSION=release/0.26
   export TORCHVISION_VERSION
 
   install_domains

.ci/docker/ubuntu/Dockerfile

@@ -64,11 +64,6 @@ ENV SCCACHE_S3_KEY_PREFIX executorch
 ENV SCCACHE_REGION us-east-1
 
 ARG TORCH_VERSION
-ARG TORCH_CHANNEL
-ARG TORCH_SPEC
-ARG TORCHAUDIO_SPEC
-ARG TORCHVISION_SPEC
-ARG TORCH_INDEX_URL
 ARG SKIP_PYTORCH
 COPY ./common/install_pytorch.sh install_pytorch.sh
 COPY ./common/utils.sh utils.sh

.ci/scripts/test_backend.sh

@@ -35,6 +35,7 @@ export PYTHON_EXECUTABLE=python
 
 # CMake options to use, in addition to the defaults.
 EXTRA_BUILD_ARGS=""
+PYTEST_RETRY_ARGS=()
 
 if [[ "$FLOW" == *qnn* ]]; then
     # Setup QNN sdk and deps - note that this is a bit hacky due to the nature of the
@@ -57,6 +58,9 @@ if [[ "$FLOW" == *vulkan* ]]; then
 fi
 
 if [[ "$FLOW" == *arm* ]]; then
+    if [[ "$SUITE" == "operators" ]]; then
+        PYTEST_RETRY_ARGS=(--reruns 2 --reruns-delay 1)
+    fi
 
     # Setup ARM deps.
     if [[ "$FLOW" == *vgf* ]]; then
@@ -95,6 +99,11 @@ GOLDEN_DIR="${ARTIFACT_DIR}/golden-artifacts"
 export GOLDEN_ARTIFACTS_DIR="${GOLDEN_DIR}"
 
 EXIT_CODE=0
-${CONDA_RUN_CMD} pytest -c /dev/null -n auto backends/test/suite/$SUITE/ -m flow_$FLOW --json-report --json-report-file="$REPORT_FILE" || EXIT_CODE=$?
+PYTEST_ARGS=(-c /dev/null -n auto)
+if [[ ${#PYTEST_RETRY_ARGS[@]} -gt 0 ]]; then
+    PYTEST_ARGS+=("${PYTEST_RETRY_ARGS[@]}")
+fi
+PYTEST_ARGS+=("backends/test/suite/$SUITE/" -m "flow_$FLOW" --json-report --json-report-file="$REPORT_FILE")
+${CONDA_RUN_CMD} pytest "${PYTEST_ARGS[@]}" || EXIT_CODE=$?
 # Generate markdown summary.
 ${CONDA_RUN_CMD} python -m executorch.backends.test.suite.generate_markdown_summary_json "$REPORT_FILE" > ${GITHUB_STEP_SUMMARY:-"step_summary.md"} --exit-code $EXIT_CODE

.ci/scripts/test_coreml_bc.sh

@@ -23,7 +23,7 @@ source "${REPO_ROOT}/.ci/scripts/utils.sh"
 # Create a conda environment with Python 3.10 for compatibility with old ET versions
 # ET 1.0.0 only supports Python >=3.10,<3.13
 CONDA_ENV_NAME="coreml_bc_test_env"
-conda create -y -n "${CONDA_ENV_NAME}" python=3.10
+conda create -y -n "${CONDA_ENV_NAME}" python=3.10 pip packaging
 
 # Use conda run to execute commands in the new environment
 CONDA_RUN="conda run --no-capture-output -n ${CONDA_ENV_NAME}"
@@ -69,7 +69,7 @@ git submodule sync --recursive
 git submodule update --init --recursive
 
 # Install executorch
-${CONDA_RUN} pip install --upgrade pip
+${CONDA_RUN} python -m pip install --upgrade pip
 ${CONDA_RUN} python install_executorch.py
 
 # Step 3: Export model
@@ -129,7 +129,7 @@ git submodule update --init --recursive
 
 # Step 5: Install current version
 echo "=== Step 5: Installing current ET version ==="
-${CONDA_RUN} pip install --upgrade pip
+${CONDA_RUN} python -m pip install --upgrade pip
 ${CONDA_RUN} python install_executorch.py
 
 # Step 6: Run the old pte file

.ci/scripts/test_huggingface_optimum_model.py

@@ -2,8 +2,10 @@
 import gc
 import logging
 import math
+import shutil
 import subprocess
 import tempfile
+import time
 from pathlib import Path
 from typing import List
 
@@ -25,6 +27,17 @@
 )
 
 
+EXPORT_RETRIES = 3
+
+
+def _clear_export_dir(model_dir):
+    for path in Path(model_dir).iterdir():
+        if path.is_dir() and not path.is_symlink():
+            shutil.rmtree(path)
+        else:
+            path.unlink()
+
+
 def cli_export(command, model_dir):
     p = Path(model_dir)
     if p.exists():
@@ -34,11 +47,19 @@ def cli_export(command, model_dir):
             raise Exception(
                 f"Existing directory {model_dir} is non-empty. Please remove it first."
             )
-    try:
-        subprocess.run(command, check=True)
-        print("Export completed successfully.")
-    except subprocess.CalledProcessError as e:
-        print(f"Export failed with error: {e}")
+
+    for attempt in range(1, EXPORT_RETRIES + 1):
+        try:
+            subprocess.run(command, check=True)
+            print("Export completed successfully.")
+            return
+        except subprocess.CalledProcessError as e:
+            print(f"Export attempt {attempt}/{EXPORT_RETRIES} failed with error: {e}")
+            if attempt == EXPORT_RETRIES:
+                raise
+            if p.exists():
+                _clear_export_dir(model_dir)
+            time.sleep(attempt * 10)
 
 
 def check_causal_lm_output_quality(

.ci/scripts/test_lora.sh

@@ -33,6 +33,24 @@ cleanup_files() {
   rm result*.txt
 }
 
+matches_base_response_prefix() {
+  local output_file="$1"
+  python - "$output_file" <<'PY'
+import pathlib
+import re
+import sys
+
+text = pathlib.Path(sys.argv[1]).read_text()
+pattern = re.compile(
+    r"^<\|im_start\|>user Calculate 15% of 80\?<\|im_end\|><\|im_start\|>assistant:\n"
+    r"(?:<think>\n)+"
+    r"Okay, so I need to calculate 15% of 80\.",
+    re.MULTILINE,
+)
+sys.exit(0 if pattern.match(text) else 1)
+PY
+}
+
 # Hosting lora adapter in personal repo for now.
 python -m pip install -q huggingface_hub
 HF_ADAPTER_REPO="lucylq/qwen3_06B_lora_math"
@@ -142,6 +160,13 @@ To calculate 15% of 80, we can multiply 80 by 15/100.
 So, 15% of 80 is equal to (80 * 15) / 100 = 1200 / 100 = 12.
 #### 12
 The answer is: 12<|im_end|>"
+EXPECTED_QUANT_LORA_ALTERNATE_PREFIX="
+<|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant
+To calculate 15% of 80, we can multiply 80 by 15/100.
+80 * 15/100 = 12.
+So, 15% of 80 is 12.
+#### 12
+The answer is: 12<|im_end|>"
 
 # Export Quantized PTE, PTD file, no LoRA.
 # override base.lora_config=null to avoid creating a lora model
@@ -186,7 +211,7 @@ cmake-out/examples/models/llama/llama_main --model_path=qwen_q.pte --data_paths=
 NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 RESULT=$(cat result.txt)
-if [[ "${RESULT}" == "${EXPECTED_QUANT_PREFIX}"* ]]; then
+if matches_base_response_prefix result.txt; then
   echo "Expected result prefix: ${EXPECTED_QUANT_PREFIX}"
   echo "Actual result: ${RESULT}"
   echo "Test 3: Success"
@@ -207,12 +232,13 @@ NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 
 RESULT=$(cat result.txt)
-if [[ "${RESULT}" == "${EXPECTED_QUANT_LORA_PREFIX}"* ]]; then
+if [[ "${RESULT}" == "${EXPECTED_QUANT_LORA_PREFIX}"* ]] || [[ "${RESULT}" == "${EXPECTED_QUANT_LORA_ALTERNATE_PREFIX}"* ]]; then
   echo "Expected result prefix: ${EXPECTED_QUANT_LORA_PREFIX}"
   echo "Actual result: ${RESULT}"
   echo "Test 4: Success"
 else
   echo "Expected result prefix: ${EXPECTED_QUANT_LORA_PREFIX}"
+  echo "Alternate expected result prefix: ${EXPECTED_QUANT_LORA_ALTERNATE_PREFIX}"
   echo "Actual result: ${RESULT}"
   echo "Test 4: Failure; results not the same"
   cleanup_files

.ci/scripts/test_lora_multimethod.sh

@@ -33,6 +33,24 @@ cleanup_files() {
   rm -f result*.txt
 }
 
+matches_base_response_prefix() {
+  local output_file="$1"
+  python - "$output_file" <<'PY'
+import pathlib
+import re
+import sys
+
+text = pathlib.Path(sys.argv[1]).read_text()
+pattern = re.compile(
+    r"^<\|im_start\|>user Calculate 15% of 80\?<\|im_end\|><\|im_start\|>assistant:\n"
+    r"(?:<think>\n)+"
+    r"Okay, so I need to calculate 15% of 80\.",
+    re.MULTILINE,
+)
+sys.exit(0 if pattern.match(text) else 1)
+PY
+}
+
 # Download LoRA adapter.
 python -m pip install -q huggingface_hub
 HF_ADAPTER_REPO="lucylq/qwen3_06B_lora_math"
@@ -107,7 +125,7 @@ NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 
 RESULT=$(cat result_base.txt)
-if [[ "${RESULT}" == "${EXPECTED_BASE_PREFIX}"* ]]; then
+if matches_base_response_prefix result_base.txt; then
   echo "Test 2 (base_forward): Success"
 else
   echo "Test 2 (base_forward): Failure"

.ci/scripts/test_model_e2e.sh

@@ -258,12 +258,21 @@ fi
 if [ "$AUDIO_URL" != "" ]; then
   curl -L $AUDIO_URL -o ${MODEL_DIR}/$AUDIO_FILE
 elif [[ "$MODEL_NAME" == *whisper* ]] || [ "$MODEL_NAME" = "voxtral_realtime" ]; then
-  conda install -y -c conda-forge "ffmpeg<8"
+  if ! command -v ffmpeg >/dev/null; then
+    if [ "$(uname -s)" = "Linux" ] && command -v apt-get >/dev/null; then
+      if [ "$(id -u)" -eq 0 ]; then
+        apt-get update
+        apt-get install -y --no-install-recommends ffmpeg
+      else
+        sudo apt-get update
+        sudo apt-get install -y --no-install-recommends ffmpeg
+      fi
+    else
+      conda install -y -c conda-forge ffmpeg
+    fi
+  fi
   pip install datasets soundfile
-  # We pushd'd into EXECUTORCH_ROOT above, so torch_pin is importable here.
-  TORCHCODEC_PKG=$(python -c "from torch_pin import torchcodec_spec; print(torchcodec_spec())")
-  TORCHCODEC_INDEX=$(python -c "from torch_pin import torch_index_url_base; print(torch_index_url_base())")
-  pip install "$TORCHCODEC_PKG" --extra-index-url "${TORCHCODEC_INDEX}/cpu"
+  pip install torchcodec==0.11.0 --extra-index-url https://download.pytorch.org/whl/test/cpu
   python -c "from datasets import load_dataset;import soundfile as sf;sample = load_dataset('distil-whisper/librispeech_long', 'clean', split='validation')[0]['audio'];sf.write('${MODEL_DIR}/$AUDIO_FILE', sample['array'][:sample['sampling_rate']*30], sample['sampling_rate'])"
 fi