Add iOS support to mobile_back_litert (CPU + Metal), upgrade to LiteRT 2.2.0

.bazelrc

@@ -26,6 +26,23 @@ build --per_file_copt=mobile_back_tflite/cpp/backend_tflite/llm_pipeline.cc@-fpe
 # tools resolved at fetch time still work.
 build --incompatible_strict_action_env
 
+# Abseil does not survive clang's layering check when it is built for the exec
+# configuration, i.e. as part of a host tool such as protoc:
+#
+#   absl/status/statusor.h:58:10: error: module com_google_absl//absl/status:
+#   statusor does not depend on a module exporting 'absl/types/span.h'
+#
+# The dependency is declared -- //absl/types:span is in statusor's deps and its
+# .cppmap is on the command line -- so this is the runner's clang disagreeing
+# with Abseil's own module layering, not a missing edge we could add. It is a
+# hygiene diagnostic over third-party code that produces no artifact, so it is
+# turned off rather than worked around.
+#
+# It went unnoticed until the LiteRT 2.2.0 pin move: the action was a remote
+# cache hit on every previous green run, and only the cold cache the new pins
+# forced actually compiled it. Target configurations keep the check.
+build --host_features=-layering_check
+
 # This flag is required by tensorflow
 common --experimental_repo_remote_exec
 

.github/workflows/ios-build-test-macos.yml

@@ -41,9 +41,15 @@ jobs:
           - backend: "tflite"
             with_apple: 0
             with_tflite: 1
+            with_litert: 0
           - backend: "apple"
             with_apple: 1
             with_tflite: 0
+            with_litert: 0
+          - backend: "litert"
+            with_apple: 0
+            with_tflite: 0
+            with_litert: 1
     env:
       PERF_TEST: true
       BAZEL_OUTPUT_ROOT_ARG: "--output_user_root=/tmp/bazel_output"
@@ -55,6 +61,7 @@ jobs:
       FLUTTER_IOS_TEST_PACKAGE: ios_tests-${{ matrix.backend }}-${{ needs.computed.outputs.build_number }}.zip
       WITH_APPLE: ${{ matrix.with_apple }}
       WITH_TFLITE: ${{ matrix.with_tflite }}
+      WITH_LITERT: ${{ matrix.with_litert }}
       WITH_PIXEL: 0
       WITH_MEDIATEK: 0
       WITH_QTI: 0
@@ -229,6 +236,9 @@ jobs:
     runs-on: ubuntu-22.04
     permissions:
       contents: read
+    # The litert row runs the llm-* benchmarks, whose go-button check validates
+    # every delegate choice's model file (two ~1.2 GB LLM exports), so it needs
+    # the same headroom as the Android unified job.
     timeout-minutes: 135
     # BrowserStack allows only 2 parallel sessions on this account, and a
     # superseded run's device tests starve the current one -- this job has
@@ -254,6 +264,8 @@ jobs:
             device: "iPhone 16 Pro-18"
           - backend: "apple"
             device: "iPhone 16 Pro-18"
+          - backend: "litert"
+            device: "iPhone 16 Pro-18"
     env:
       TEST_PACKAGE_NAME: ios_tests-${{ matrix.backend }}-${{ needs.computed.outputs.build_number }}.zip
       BROWSERSTACK_LOGS_DIR: /tmp/browserstack-device-logs
@@ -282,11 +294,19 @@ jobs:
           BROWSERSTACK_DEVICES: >-
             ["${{ matrix.device }}"]
         with:
-          timeout_minutes: 60
+          # This is the timeout that actually fires: the job-level one never
+          # gets a chance, because the step waits out retry_wait_seconds even
+          # when it will not retry (retry_on_exit_code only covers a trigger
+          # rejection). Keep it below the job cap minus retry_wait_seconds.
+          # BrowserStack reports a build as "running" while it is still queued
+          # for a device, so this budget covers queue time as well as the run.
+          timeout_minutes: 80
           # Exit 9 is a BrowserStack trigger failure, which is what
           # BROWSERSTACK_ALL_PARALLELS_IN_USE produces when the account's
           # session budget is busy - retrying it just waits for a free slot.
-          # Every other failure still fails fast.
+          # Every other failure still fails fast. A rejection returns
+          # immediately, but each retry still costs retry_wait_seconds, so a
+          # long queue can burn into the job cap above.
           max_attempts: 8
           retry_wait_seconds: 300
           retry_on_exit_code: 9

.github/workflows/scripts/browserstack-app-automate.sh

@@ -125,16 +125,27 @@
   if [[ -n "$test_id" && "$test_id" != "null" && -n "$device_log_url" && "$device_log_url" != "null" ]]; then
     echo "Found last test case $test_id with device log URL"
 
-    # Download device logs using the extracted URL
+    # Download device logs using the extracted URL. -f so an HTTP error is a
+    # failure instead of an error page written to the log, and retries because
+    # the log is not always served the instant the session ends.
     local log_file="$LOGS_DIR/${test_id}.log"
     echo "Downloading device log to $log_file"
-    curl -s -u "$CREDENTIALS" -X GET "$device_log_url" -o "$log_file"
-
-    if [ -f "$log_file" ]; then
-      echo "Device logs downloaded successfully to $log_file"
+    if curl -sf --retry 3 --retry-delay 5 -u "$CREDENTIALS" -X GET \
+        "$device_log_url" -o "$log_file" && [ -s "$log_file" ]; then
+      echo "Device logs downloaded successfully to $log_file ($(wc -c <"$log_file") bytes)"
     else
-      echo "Failed to download device logs for test case $test_id"
+      # Test -s, not -f: a zero-byte log is what this used to leave behind, and
+      # it reads as "the run produced no output" rather than "the download
+      # failed". Leave a note in its place so the artifact explains itself --
+      # if-no-files-found on the upload step only catches a missing file.
+      echo "Failed to download device logs for test case $test_id" \
+        | tee "$log_file"
+      echo "  session: $session_id" >> "$log_file"
+      echo "  url: $device_log_url" >> "$log_file"
     fi
+  else
+    echo "No device log URL in the session response for build $build_id" \
+      | tee "$LOGS_DIR/no-device-log-${session_id}.txt"
   fi
 }
 

WORKSPACE

@@ -30,7 +30,7 @@ http_archive(
 # rules_python that XLA now brings in: rules_apple's plisttool is generated
 # from a bootstrap template containing %interpreter_args%, which the older
 # rules leave unsubstituted, so it fails to parse as Python. These are the
-# same versions LiteRT 2.1.5 pins for this dependency set.
+# same versions LiteRT 2.2.0 pins for this dependency set.
 http_archive(
     name = "build_bazel_rules_apple",
     sha256 = "a78f26c22ac8d6e3f3fcaad50eace4d9c767688bd7254b75bdf4a6735b299f6a",
@@ -51,6 +51,23 @@ http_archive(
     url = "https://github.com/bazelbuild/apple_support/releases/download/1.23.1/apple_support.1.23.1.tar.gz",
 )
 
+# The rules_cc rules_python's py_repositories() would bring in anyway, patched.
+# Declared here so it is fetched with the patch rather than without it; the
+# declaration below it is a maybe(), so ours wins.
+#
+# Moving TensorFlow to LiteRT 2.2.0's commit pulled this chain -- XLA, then
+# rules_ml_toolchain, then rules_python -- forward to a rules_cc whose
+# use_cc_toolchain() is mandatory. That breaks the Android build during
+# analysis. See patches/rules_cc_optional_cc_toolchain.patch for the detail.
+http_archive(
+    name = "rules_cc",
+    patch_args = ["-p1"],
+    patches = ["//patches:rules_cc_optional_cc_toolchain.patch"],
+    sha256 = "b8b918a85f9144c01f6cfe0f45e4f2838c7413961a8ff23bc0c6cdf8bb07a3b6",
+    strip_prefix = "rules_cc-0.1.5",
+    url = "https://github.com/bazelbuild/rules_cc/releases/download/0.1.5/rules_cc-0.1.5.tar.gz",
+)
+
 http_archive(
     name = "bazel_features",
     sha256 = "c26b4e69cf02fea24511a108d158188b9d8174426311aac59ce803a78d107648",
@@ -131,21 +148,21 @@ http_archive(
     patch_args = ["-p1"],
     patches = [
         # Add patches for adding png in tflite evaluation code
+        # Channel detection and the unsigned-char decode buffer, formerly
+        # png-with-number-of-channels-detected.patch and use_unsigned_char.patch,
+        # are folded into this one.
         "//:flutter/third_party/enable-png-in-tensorflow-lite-tools-evaluation.patch",
-        "//:flutter/third_party/png-with-number-of-channels-detected.patch",
-        "//:flutter/third_party/use_unsigned_char.patch",
         # Fix tensorflow not being able to read image files on Windows
         "//:flutter/third_party/tensorflow-fix-file-opening-mode-for-Windows.patch",
+        "//:patches/litert_logistic_fp16_msvc.patch",
         "//:patches/litert-internal-visibility.diff",
         # Fix for LiteRT crashing on close when using OpenCL accelerator
         "//:patches/custom_buffer_teardown.patch",
-        # CoreML delegate calls RepeatedField::resize, which does not exist
-        "//:patches/litert_coreml_repeatedfield_resize.patch",
     ],
-    sha256 = "7d0313c4851deb18af6f5f2dbc002bf01293583b87b819b0949ee33dcfe2d91b",
-    strip_prefix = "LiteRT-2.1.5",
+    sha256 = "6d2ce16738199adc5a3cdde76c3c6a6dac636d3b52a1d7790ea524fb0d59f7fc",
+    strip_prefix = "LiteRT-2.2.0",
     urls = [
-        "https://github.com/google-ai-edge/LiteRT/archive/v2.1.5.tar.gz",
+        "https://github.com/google-ai-edge/LiteRT/archive/v2.2.0.tar.gz",
     ],
 )
 
@@ -161,13 +178,17 @@ tensorflow_source_repo(
     patches = [
         "//:flutter/third_party/tf-eigen.patch",
         "//patches:tf_coreml_repeatedfield_resize.patch",
+        "//patches:tf_logistic_fp16_msvc.patch",
         "//patches:tf_nnapi_no_mmap_sharing.patch",
         "//patches:tf_portable_no_onednn_env_vars.patch",
     ] + PATCH_FILE,
-    sha256 = "879cf25692d50c60315a4dd3929dccd923d4c44a2c4b95ebb483666d2c16a22a",
-    strip_prefix = "tensorflow-6d40c20cdfe385746c31da6227b95722f5ece342",
+    # The commit LiteRT 2.2.0 pins. LiteRT's GPU delegate needs
+    # @com_google_absl//absl/status:status_macros, which only exists in the
+    # Abseil this TensorFlow's XLA brings in, so the two move together.
+    sha256 = "c3c552414ab2e59e72511a21c1df566346a7c8f160909325edec6d1ff403d69d",
+    strip_prefix = "tensorflow-bcdab1a62e138c8f8784a7477c0be8af6dd0bd0a",
     urls = [
-        "https://github.com/tensorflow/tensorflow/archive/6d40c20cdfe385746c31da6227b95722f5ece342.tar.gz",
+        "https://github.com/tensorflow/tensorflow/archive/bcdab1a62e138c8f8784a7477c0be8af6dd0bd0a.tar.gz",
     ],
 )
 

flutter/assets/tasks.pbtxt

@@ -417,7 +417,13 @@ task {
     quick {
       min_query_count: 10
       min_duration: 10
-      max_duration: 40
+      # 40s could never satisfy min_query_count on a 1B model: measured on an
+      # iPhone 16 Pro, llm-1b runs about 10.1s per query on the Metal delegate
+      # (24.3 tok/s), so ten queries need ~101s and the run ended invalid --
+      # isMinQueryMet false -- on both the CPU and the GPU delegate. A run
+      # stops as soon as min_duration AND min_query_count are both met, so
+      # this ceiling only costs time on devices that were failing anyway.
+      max_duration: 150
     }
     rapid {
       min_query_count: 6
@@ -477,7 +483,13 @@ task {
     quick {
       min_query_count: 10
       min_duration: 10
-      max_duration: 40
+      # 40s could never satisfy min_query_count on a 1B model: measured on an
+      # iPhone 16 Pro, llm-1b runs about 10.1s per query on the Metal delegate
+      # (24.3 tok/s), so ten queries need ~101s and the run ended invalid --
+      # isMinQueryMet false -- on both the CPU and the GPU delegate. A run
+      # stops as soon as min_duration AND min_query_count are both met, so
+      # this ceiling only costs time on devices that were failing anyway.
+      max_duration: 150
     }
     rapid {
       min_query_count: 6
@@ -657,7 +669,13 @@ task {
     quick {
       min_query_count: 10
       min_duration: 10
-      max_duration: 40
+      # 40s could never satisfy min_query_count on a 1B model: measured on an
+      # iPhone 16 Pro, llm-1b runs about 10.1s per query on the Metal delegate
+      # (24.3 tok/s), so ten queries need ~101s and the run ended invalid --
+      # isMinQueryMet false -- on both the CPU and the GPU delegate. A run
+      # stops as soon as min_duration AND min_query_count are both met, so
+      # this ceiling only costs time on devices that were failing anyway.
+      max_duration: 150
     }
     rapid {
       min_query_count: 6
@@ -717,7 +735,13 @@ task {
     quick {
       min_query_count: 10
       min_duration: 10
-      max_duration: 40
+      # 40s could never satisfy min_query_count on a 1B model: measured on an
+      # iPhone 16 Pro, llm-1b runs about 10.1s per query on the Metal delegate
+      # (24.3 tok/s), so ten queries need ~101s and the run ended invalid --
+      # isMinQueryMet false -- on both the CPU and the GPU delegate. A run
+      # stops as soon as min_duration AND min_query_count are both met, so
+      # this ceiling only costs time on devices that were failing anyway.
+      max_duration: 150
     }
     rapid {
       min_query_count: 6
@@ -777,7 +801,13 @@ task {
     quick {
       min_query_count: 10
       min_duration: 10
-      max_duration: 40
+      # 40s could never satisfy min_query_count on a 1B model: measured on an
+      # iPhone 16 Pro, llm-1b runs about 10.1s per query on the Metal delegate
+      # (24.3 tok/s), so ten queries need ~101s and the run ended invalid --
+      # isMinQueryMet false -- on both the CPU and the GPU delegate. A run
+      # stops as soon as min_duration AND min_query_count are both met, so
+      # this ceiling only costs time on devices that were failing anyway.
+      max_duration: 150
     }
     rapid {
       min_query_count: 6
@@ -837,7 +867,13 @@ task {
     quick {
       min_query_count: 10
       min_duration: 10
-      max_duration: 40
+      # 40s could never satisfy min_query_count on a 1B model: measured on an
+      # iPhone 16 Pro, llm-1b runs about 10.1s per query on the Metal delegate
+      # (24.3 tok/s), so ten queries need ~101s and the run ended invalid --
+      # isMinQueryMet false -- on both the CPU and the GPU delegate. A run
+      # stops as soon as min_duration AND min_query_count are both met, so
+      # this ceiling only costs time on devices that were failing anyway.
+      max_duration: 150
     }
     rapid {
       min_query_count: 6

flutter/cpp/datasets/ifeval_utils/BUILD

@@ -40,3 +40,13 @@ cc_library(
         "@oleander_stemming_library",
     ],
 )
+
+cc_test(
+    name = "common_test",
+    srcs = ["common_test.cc"],
+    linkstatic = 1,
+    deps = [
+        ":ifeval_utils",
+        "@com_google_googletest//:gtest",
+    ],
+)

flutter/cpp/datasets/ifeval_utils/common.h

@@ -54,7 +54,10 @@ inline bool contains_string(const std::string& text,
 inline bool ends_with(const std::string& s, const std::string& suf,
                       unsigned threshold) {
   if (s.size() < suf.size()) return false;
-  std::string a = tolower(s.substr(s.size() - (suf.size() + threshold)));
+  // The window may be longer than the response itself, in which case it is the
+  // whole response; subtracting unclamped would wrap and make substr throw.
+  const std::size_t window = std::min(s.size(), suf.size() + threshold);
+  std::string a = tolower(s.substr(s.size() - window));
   std::string b = tolower(suf);
   return threshold == 0 ? a == b : contains_string(a, b);
 }
@@ -142,8 +145,10 @@ inline std::string remove_font_modifiers(const std::string& s) {
     }
 
     // skip emphasis/strong/strike/escape chars as long as they're not preceeded
-    // by an escape character
-    if ((c == '*' || c == '_' || c == '~' || c == '\\') && s[i - 1] != '\\')
+    // by an escape character. The first character has nothing before it, so it
+    // is never escaped -- without the i == 0 guard this reads s[SIZE_MAX].
+    if ((c == '*' || c == '_' || c == '~' || c == '\\') &&
+        (i == 0 || s[i - 1] != '\\'))
       continue;
 
     // remove heading markers (#) at line starts