Update Gemma 4 benchmarks with static INT2 SME2 kernels

content/learning-paths/mobile-graphics-and-gaming/gemma4-kleidiai-sme2/_index.md

@@ -1,74 +1,28 @@
 ---
-title: Benchmark Gemma 4 LiteRT-LM prefill performance with KleidiAI and SME2 on macOS
-description: Benchmark Gemma 4 prefill performance on macOS by comparing baseline and SME2-optimized XNNPACK paths in LiteRT-LM with KleidiAI.
+title: Benchmark Gemma 4 CPU inference with KleidiAI and SME2
+description: Benchmark Gemma 4 prefill and decode on macOS with LiteRT-LM and KleidiAI SME2 kernels, and compare results from two devices.
 
-minutes_to_complete: 45
+minutes_to_complete: 90
 
-draft: true
+draft: false
 cascade:
-    draft: true
+    draft: false
 
-who_is_this_for: This is an advanced topic for software developers and performance engineers who want a reproducible Gemma 4 prefill benchmark workflow using LiteRT-LM, KleidiAI, and XNNPACK on macOS.
+who_is_this_for: Software developers and performance engineers evaluating Gemma 4 CPU inference with LiteRT-LM, KleidiAI, and XNNPACK on SME2 hardware.
 
 learning_objectives:
-    - Create a workspace with pinned LiteRT-LM and KleidiAI versions.
-    - Create upstream-optimized and historical-baseline XNNPACK worktrees.
-    - Download a LiteRT-LM-compatible Gemma 4 model from Hugging Face.
-    - Compare baseline and upstream SME2 benchmark results.
+    - Build LiteRT-LM with baseline and optimized XNNPACK kernels
+    - Download and verify the Gemma 4 E2B LiteRT-LM model
+    - Measure prefill and decode throughput at one and four CPU threads
+    - Compare throughput and verify the accelerated operator
 
 prerequisites:
-    - An SME2 device (macOS on Apple M4 is used in this Learning Path for demonstrative purposes)
+    - An [SME2-capable device](/learning-paths/cross-platform/multiplying-matrices-with-sme2/1-get-started/#devices), such as an M4-family Mac
     - Git, Homebrew, and Xcode Command Line Tools
     - At least 25 GB of free disk space for model files and local builds
-
-# START generated_summary_faq
-generated_summary_faq:
-  template_version: summary-faq-v3
-  generated_at: '2026-09-14T20:55:19Z'
-  generator: ai
-  ai_assisted: true
-  ai_review_required: true
-  model: gpt-5
-  prompt_template: summary-faq-v3
-  source_hash: 87d7867854c80eb5b1a520c8958d4251d51e5e33c98f4d7b1890431ca82cfe8b
-  summary_generated_at: '2026-09-14T20:55:19Z'
-  summary_source_hash: 87d7867854c80eb5b1a520c8958d4251d51e5e33c98f4d7b1890431ca82cfe8b
-  faq_generated_at: '2026-09-14T20:55:19Z'
-  faq_source_hash: 87d7867854c80eb5b1a520c8958d4251d51e5e33c98f4d7b1890431ca82cfe8b
-  summary: >-
-    You'll benchmark Gemma 4 prefill performance on macOS by comparing an upstream SME2-optimized
-    XNNPACK path with a historical baseline in LiteRT-LM and KleidiAI. First, you'll set up a workspace and
-    verify SME and SME2 availability. Then, you'll install Bazelisk and the Hugging Face Hub CLI, and download
-    a LiteRT-LM-compatible Gemma 4 E2B `.litertlm` model. You'll build both XNNPACK variants with
-    identical settings, then compare three-iteration results at one and four CPU threads.
-  faqs:
-  - question: How do I check that my device supports SME2 before running the benchmarks?
-    answer: >-
-      Run `uname -m` and the `sysctl` checks for `hw.optional.arm.FEAT_SME` and
-      `hw.optional.arm.FEAT_SME2`. Proceed when the architecture is `arm64` and both feature flags
-      report `1`.
-  - question: Which Gemma 4 model should I download?
-    answer: >-
-     Download the Gemma 4 E2B `.litertlm` model from Hugging
-      Face. Place it in the shared model directory at `$HOME/gemma4-prefill-bench/models`.
-  - question: How do I confirm that Bazel is the version expected by LiteRT-LM?
-    answer: >-
-      From the LiteRT-LM repository, check `.bazelversion` and run `bazelisk version`. Confirm that the Bazel version is `7.6.1`.
-  - question: Why should I create separate baseline and upstream XNNPACK trees?
-    answer: >-
-      You'll use separate trees to build and benchmark both variants with identical settings while
-      recording the exact repository revisions. This enables a clean side-by-side comparison through
-      the same LiteRT-LM entrypoint.
-  - question: What does the benchmark function run, and where are the results saved?
-    answer: >-
-      When you run the benchmark function, it builds the selected XNNPACK variant and runs
-      three-iteration prefill benchmarks at one and four CPU threads. Results are written under the
-      workspace results directory, so you can compare baseline and SME2-optimized runs.
-# END generated_summary_faq
-
 author: Annie Tallund
 
-generate_summary_faq: false
+generate_summary_faq: true
 rerun_summary: false
 rerun_faqs: false
 
@@ -91,10 +45,6 @@ further_reading:
         title: Arm Scalable Matrix Extension introduction, part 1
         link: https://developer.arm.com/community/arm-community-blogs/b/architectures-and-processors-blog/posts/arm-scalable-matrix-extension-introduction
         type: blog
-    - resource:
-        title: Arm Scalable Matrix Extension instructions, part 2
-        link: https://developer.arm.com/community/arm-community-blogs/b/architectures-and-processors-blog/posts/arm-scalable-matrix-extension-introduction-p2
-        type: blog
     - resource:
         title: Arm SME2 introduction, part 4
         link: https://developer.arm.com/community/arm-community-blogs/b/architectures-and-processors-blog/posts/part4-arm-sme2-introduction

content/learning-paths/mobile-graphics-and-gaming/gemma4-kleidiai-sme2/download_model.md

@@ -1,40 +1,36 @@
 ---
 title: Install prerequisites on macOS and download Gemma 4
-description: Install Bazelisk and the Hugging Face Hub CLI on macOS, then download the Gemma 4 LiteRT-LM model artifact for benchmarking.
+description: Install the macOS build tools and verify the Gemma 4 E2B LiteRT-LM model used for both XNNPACK variants.
 weight: 3
 
 ### FIXED, DO NOT MODIFY
 layout: learningpathall
 ---
 
-## Install Bazelisk and the Hugging Face Hub CLI
+## Install prerequisites on macOS
 
-Install the prerequisite software for the benchmark.
-
-Install Bazelisk:
+Install Xcode Command Line Tools if they are not already available:
 
 ```bash
-brew install bazelisk
+xcode-select --install
 ```
 
-Install the Hugging Face Hub CLI:
+Install Bazelisk:
 
 ```bash
-python3 -m pip install -U huggingface_hub
+brew install bazelisk
 ```
 
-Confirm that LiteRT-LM pins Bazel `7.6.1`:
+Install the Hugging Face Hub CLI in a virtual environment:
 
 ```bash
-cd $HOME/gemma4-prefill-bench/LiteRT-LM
-cat .bazelversion
-bazelisk version
+python3 -m venv "$HOME/gemma4-prefill-bench/.venv"
+source "$HOME/gemma4-prefill-bench/.venv/bin/activate"
+python3 -m pip install -U huggingface_hub
 ```
 
 ## Download Gemma 4 from Hugging Face
 
-With the prerequisites installed, download the Gemma 4 E2B `.litertlm` model from Hugging Face.
-
 Create a shared model directory in the workspace:
 
 ```bash
@@ -47,18 +43,17 @@ Download the CPU-compatible Gemma 4 E2B LiteRT-LM artifact:
 ```bash
 hf download litert-community/gemma-4-E2B-it-litert-lm \
   gemma-4-E2B-it.litertlm \
+  --revision 6b78abd019e61a1ca4cbe3b212d2c9ce8ff38a94 \
   --local-dir models/gemma-4-E2B-it-litert-lm
 ```
 
-The model repository is public. If Hugging Face asks for credentials, run
-`hf auth login` and repeat the download.
+If Hugging Face asks for credentials, run `hf auth login` and repeat the download.
 
 Verify the downloaded file:
 
 ```bash
 shasum -a 256 \
   models/gemma-4-E2B-it-litert-lm/gemma-4-E2B-it.litertlm
-ls -lh models/gemma-4-E2B-it-litert-lm/gemma-4-E2B-it.litertlm
 ```
 
 The expected SHA-256 checksum is:
@@ -68,13 +63,8 @@ The expected SHA-256 checksum is:
 ```
 
 {{% notice Note %}}
-The file is about 2.6 GB. LiteRT-LM requires the `.litertlm` artifact. A
-Transformers repository containing only `safetensors` files isn't a direct
-replacement.
+Download the `.litertlm` artifact used by LiteRT-LM; the file is about 2.6 GB.
 {{% /notice %}}
 
-## What you've accomplished and what's next
-
-You've installed prerequisite software on your macOS device and downloaded Gemma 4 from Hugging Face.
-
-Next, you'll build each XNNPACK variant and run the benchmark.
+You've installed the build tools and verified the shared model artifact.
+Next, build each XNNPACK variant and run the benchmark.