cuda : widen QSA selected-key block ids to uint32 and lift the depth bound to 1M

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -13,7 +13,7 @@ body:
         If the compilation succeeds with ccache disabled you should be able to permanently fix the issue
         by clearing `~/.cache/ccache` (on Linux).
 
-        Please fill out this template yourself, copypasting language model outputs is [strictly prohibited](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
+        AI assistance is allowed in this fork. Disclose how it was used, verify the information in your report, and take responsibility for its accuracy. See [our AI usage policy](https://github.com/halo-box/strix-llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
   - type: textarea
     id: commit
     attributes:

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -13,7 +13,7 @@ body:
         please reproduce your issue using one of the examples/binaries in this repository.
         The `llama-completion` binary can be used for simple and reproducible model inference.
 
-        Please fill out this template yourself, copypasting language model outputs is [strictly prohibited](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
+        AI assistance is allowed in this fork. Disclose how it was used, verify the information in your report, and take responsibility for its accuracy. See [our AI usage policy](https://github.com/halo-box/strix-llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
   - type: textarea
     id: version
     attributes:

.github/ISSUE_TEMPLATE/019-bug-misc.yml

@@ -11,7 +11,7 @@ body:
         If you encountered the issue while using an external UI (e.g. ollama),
         please reproduce your issue using one of the examples/binaries in this repository.
 
-        Please fill out this template yourself, copypasting language model outputs is [strictly prohibited](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
+        AI assistance is allowed in this fork. Disclose how it was used, verify the information in your report, and take responsibility for its accuracy. See [our AI usage policy](https://github.com/halo-box/strix-llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
   - type: textarea
     id: version
     attributes:

.github/ISSUE_TEMPLATE/020-enhancement.yml

@@ -8,7 +8,7 @@ body:
       value: |
         [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggml-org/llama.cpp/discussions/categories/ideas)
 
-        Please fill out this template yourself, copypasting language model outputs is [strictly prohibited](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
+        AI assistance is allowed in this fork. Disclose how it was used, verify the information in your report, and take responsibility for its accuracy. See [our AI usage policy](https://github.com/halo-box/strix-llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
 
   - type: checkboxes
     id: prerequisites

.github/ISSUE_TEMPLATE/030-research.yml

@@ -8,7 +8,7 @@ body:
       value: |
         Don't forget to check for any [duplicate research issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
 
-        Please fill out this template yourself, copypasting language model outputs is [strictly prohibited](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
+        AI assistance is allowed in this fork. Disclose how it was used, verify the information in your report, and take responsibility for its accuracy. See [our AI usage policy](https://github.com/halo-box/strix-llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
 
   - type: checkboxes
     id: research-stage

.github/ISSUE_TEMPLATE/040-refactor.yml

@@ -9,7 +9,7 @@ body:
         Don't forget to [check for existing refactor issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
         Also you may want to check [Pull request refactor label as well](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
 
-        Please fill out this template yourself, copypasting language model outputs is [strictly prohibited](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
+        AI assistance is allowed in this fork. Disclose how it was used, verify the information in your report, and take responsibility for its accuracy. See [our AI usage policy](https://github.com/halo-box/strix-llama.cpp/blob/master/CONTRIBUTING.md#ai-usage-policy).
 
   - type: textarea
     id: background-description

.github/actions/test-llamacpp-build/action.yml

@@ -0,0 +1,238 @@
+# Same as lemonade-sdk/llamacpp-rocm (MIT).
+name: 'Test Llama.cpp Build'
+description: 'Download and test llama.cpp build artifacts'
+
+inputs:
+  os_type:
+    description: 'Operating system type (Windows or Linux)'
+    required: true
+  gfx_target:
+    description: 'GPU target (e.g., gfx1151, gfx1150)'
+    required: true
+  artifact_name:
+    description: 'Full artifact name to download'
+    required: true
+
+runs:
+  using: "composite"
+  steps:
+    - name: Download build artifacts
+      uses: actions/download-artifact@v4
+      with:
+        name: ${{ inputs.artifact_name }}
+        path: llama-binaries
+
+    - name: Download test model (Windows)
+      if: inputs.os_type == 'Windows'
+      shell: powershell
+      run: |
+        $modelUrl = "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-Q4_0.gguf?download=true"
+        $modelPath = "Qwen3-0.6B-Q4_0.gguf"
+
+        Write-Host "Downloading test model from: $modelUrl"
+        curl.exe -L -o $modelPath $modelUrl
+
+        # Verify download
+        if (Test-Path $modelPath) {
+          $fileSize = (Get-Item $modelPath).Length
+          Write-Host "Model downloaded successfully. Size: $fileSize bytes"
+        } else {
+          Write-Error "Failed to download model"
+          exit 1
+        }
+
+    - name: Download test model (Linux)
+      if: inputs.os_type == 'Linux'
+      shell: bash
+      run: |
+        model_url="https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-Q4_0.gguf?download=true"
+        model_path="Qwen3-0.6B-Q4_0.gguf"
+
+        echo "Downloading test model from: $model_url"
+        curl -L -o "$model_path" "$model_url"
+
+        # Verify download
+        if [ -f "$model_path" ]; then
+          file_size=$(stat -c%s "$model_path")
+          echo "Model downloaded successfully. Size: $file_size bytes"
+        else
+          echo "Failed to download model"
+          exit 1
+        fi
+
+    - name: Set up library path (Linux)
+      if: inputs.os_type == 'Linux'
+      shell: bash
+      run: |
+        # Add the llama-binaries directory to LD_LIBRARY_PATH
+        echo "LD_LIBRARY_PATH=$(pwd)/llama-binaries:$LD_LIBRARY_PATH" >> $GITHUB_ENV
+
+    - name: Run llama-cli test (Windows)
+      if: inputs.os_type == 'Windows'
+      shell: powershell
+      run: |
+        $llamaCliPath = "llama-binaries\llama-cli.exe"
+        $modelPath = "Qwen3-0.6B-Q4_0.gguf"
+
+        # Verify llama-cli exists
+        if (-not (Test-Path $llamaCliPath)) {
+          Write-Error "llama-cli.exe not found at: $llamaCliPath"
+          Write-Host "Available files in llama-binaries:"
+          Get-ChildItem -Recurse "llama-binaries"
+          exit 1
+        }
+
+        Write-Host "Running llama-cli test for ${{ inputs.gfx_target }}..."
+        Write-Host "Command: $llamaCliPath -m `"$modelPath`" -ngl 99 -p `"I believe the meaning of life is`" -st -v"
+
+        # Run the command and capture output
+        try {
+          $process = Start-Process -FilePath $llamaCliPath -ArgumentList @("-m", "`"$modelPath`"", "-ngl", "99", "-p", "`"I believe the meaning of life is`"", "-st", "-v") -NoNewWindow -Wait -PassThru -RedirectStandardOutput "stdout.txt" -RedirectStandardError "stderr.txt"
+
+          # Wait for the process to complete
+          $process.WaitForExit()
+
+          # Read both stdout and stderr
+          $stdout = Get-Content "stdout.txt" -Raw -ErrorAction SilentlyContinue
+          $stderr = Get-Content "stderr.txt" -Raw -ErrorAction SilentlyContinue
+
+          # Always print the output (both stdout and stderr)
+          Write-Host "=== LLAMA-CLI STDOUT ==="
+          if ($stdout) {
+            Write-Host $stdout
+          } else {
+            Write-Host "(empty)"
+          }
+          Write-Host "=== LLAMA-CLI STDERR ==="
+          if ($stderr) {
+            Write-Host $stderr
+          } else {
+            Write-Host "(empty)"
+          }
+          Write-Host "=== END OUTPUT ==="
+          Write-Host "Process exit code: $($process.ExitCode)"
+
+          # Combine output for pattern matching (stderr often contains important CUDA/GPU info)
+          $output = $stdout + $stderr
+
+          # Clean up temp files
+          Remove-Item "stdout.txt" -ErrorAction SilentlyContinue
+          Remove-Item "stderr.txt" -ErrorAction SilentlyContinue
+        }
+        catch {
+          Write-Error "Failed to start llama-cli process: $_"
+          Write-Error "This usually indicates a missing dependency (DLL) or permission issue"
+          # Try to get more details about dependencies
+          Write-Host "Checking for dependency issues..."
+          Write-Host "Files in llama-binaries directory:"
+          Get-ChildItem "llama-binaries" | Format-Table Name, Length
+          exit 1
+        }
+
+        # Check if the process exited successfully
+        if ($process.ExitCode -ne 0) {
+          Write-Host "❌ llama-cli exited with error code: $($process.ExitCode)"
+          Write-Host "This error code typically indicates:"
+          Write-Host "  - Exit code 127: Command or library not found (missing DLL)"
+          Write-Host "  - Exit code 1: General error"
+          Write-Host "  - Other codes: Application-specific errors"
+          Write-Host ""
+          Write-Host "Please check the STDERR output above for specific error messages."
+          exit 1
+        }
+
+        # Check for expected outputs
+        $foundGpuOffload = $output -match "offloaded 29/29 layers to GPU"
+        $foundThinkTag = $output -match "</think>"
+
+        Write-Host "=== TEST RESULTS ==="
+        Write-Host "Expected 'offloaded 29/29 layers to GPU': $(if ($foundGpuOffload) { 'FOUND' } else { 'NOT FOUND' })"
+        Write-Host "Expected '</think>' tag: $(if ($foundThinkTag) { 'FOUND' } else { 'NOT FOUND' })"
+
+        if ($foundGpuOffload -and $foundThinkTag) {
+          Write-Host "✅ Test PASSED - Both expected outputs found"
+        } else {
+          Write-Host "❌ Test FAILED - Missing expected outputs"
+          exit 1
+        }
+
+    - name: Run llama-cli test (Linux)
+      if: inputs.os_type == 'Linux'
+      shell: bash
+      run: |
+        llama_cli_path="./llama-binaries/llama-cli"
+        model_path="Qwen3-0.6B-Q4_0.gguf"
+        output_file="llama_output.txt"
+
+        # Make llama-cli executable
+        chmod +x "$llama_cli_path"
+
+        # Verify llama-cli exists
+        if [ ! -f "$llama_cli_path" ]; then
+          echo "llama-cli not found at: $llama_cli_path"
+          echo "Available files in llama-binaries:"
+          find llama-binaries -type f
+          exit 1
+        fi
+
+        echo "Running llama-cli test for ${{ inputs.gfx_target }}..."
+        echo "Command: $llama_cli_path -m \"$model_path\" -ngl 99 -p \"I believe the meaning of life is\" -st -v"
+
+        # Run the command and capture output to file (more robust than variable)
+        set +e  # Don't exit on error, we want to capture the output and check exit code manually
+        "$llama_cli_path" -m "$model_path" -ngl 99 -p "I believe the meaning of life is" -st -v > "$output_file" 2>&1
+        exit_code=$?
+        set -e  # Re-enable exit on error
+
+        echo "=== LLAMA-CLI OUTPUT ==="
+        if [ -f "$output_file" ] && [ -s "$output_file" ]; then
+          cat "$output_file"
+        else
+          echo "(empty)"
+        fi
+        echo "=== END OUTPUT ==="
+        echo "Process exit code: $exit_code"
+
+        # Check if the process exited successfully
+        if [ $exit_code -ne 0 ]; then
+          echo "❌ llama-cli exited with error code: $exit_code"
+          echo "This error code typically indicates:"
+          echo "  - Exit code 127: Command or library not found (missing shared library)"
+          echo "  - Exit code 1: General error"
+          echo "  - Other codes: Application-specific errors"
+          echo ""
+          echo "Checking for missing library dependencies..."
+          ldd "$llama_cli_path" || echo "ldd command failed"
+          echo ""
+          echo "Please check the output above for specific error messages."
+          rm -f "$output_file"
+          exit 1
+        fi
+
+        # Check for expected outputs using grep -F (fixed strings, no regex interpretation)
+        # This handles special characters like angle brackets more reliably
+        found_gpu_offload=false
+        found_think_tag=false
+
+        if grep -F "offloaded 29/29 layers to GPU" "$output_file" > /dev/null 2>&1; then
+          found_gpu_offload=true
+        fi
+
+        if grep -F "</think>" "$output_file" > /dev/null 2>&1; then
+          found_think_tag=true
+        fi
+
+        echo "=== TEST RESULTS ==="
+        echo "Expected 'offloaded 29/29 layers to GPU': $(if [ "$found_gpu_offload" = true ]; then echo 'FOUND'; else echo 'NOT FOUND'; fi)"
+        echo "Expected '</think>' tag: $(if [ "$found_think_tag" = true ]; then echo 'FOUND'; else echo 'NOT FOUND'; fi)"
+
+        # Clean up output file
+        rm -f "$output_file"
+
+        if [ "$found_gpu_offload" = true ] && [ "$found_think_tag" = true ]; then
+          echo "✅ Test PASSED - Both expected outputs found"
+        else
+          echo "❌ Test FAILED - Missing expected outputs"
+          exit 1
+        fi
+

.github/pull_request_template.md

@@ -2,6 +2,37 @@
 
 <!-- Describe what this PR does and why. Be concise but complete -->
 
+## Measurements
+
+<!--
+Required for anything that can move throughput, latency, memory use, or output. Delete this section only for docs,
+comments, CI or build changes that cannot affect runtime - and say that is why.
+Full rules: CONTRIBUTING.md#benchmarking-requirements
+-->
+
+```
+Device:     Ryzen AI Max+ 395 / <mini-PC or laptop model>
+Memory:     128 GB LPDDR5X-8000
+Power:      <sustained TDP / platform profile / power governor>
+BIOS:       UMA split <n> GB
+Kernel:     <uname -r>, amdgpu params: <gttsize / ttm.pages_limit, or none>
+Backend:    Vulkan RADV, Mesa <version>   (or: ROCm <version>, HIP_LAUNCH_BLOCKING=<0|1>)
+Build:      <cmake flags>
+Baseline:   <merge-base commit sha, built and run in this same session>
+Change:     <your branch commit sha>
+Model:      <HF repo / file>, <quant>
+```
+
+<!-- Paste the raw llama-bench tables for baseline and change, with their standard-deviation columns. -->
+
+**Baseline:**
+
+**After:**
+
+<!-- Output-equivalence evidence: test-backend-ops for the backend you touched, and/or llama-perplexity before/after. -->
+
+**Correctness:**
+
 ## Additional information
 
 <!-- You can provide more details and link related discussions here. Delete this section if not applicable -->
@@ -10,7 +41,14 @@
 
 <!-- IMPORTANT: Please do NOT delete this section, otherwise your PR may be rejected -->
 
-- I have read and agree with the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md)
-- AI usage disclosure: <!-- mention: YES / NO - if yes, describe how AI was used -->
+- I have read and agree with the [contributing guidelines](CONTRIBUTING.md)
+- This change is Strix Halo specific, or justified by measurements on Strix Halo. General llama.cpp improvements
+  belong in [halo-box/llama.cpp](https://github.com/halo-box/llama.cpp) instead
+- AI usage disclosure: <!-- mention: NO / ASSISTED / AGENT-AUTHORED - and describe how AI was used -->
+- What was NOT verified: <!-- untested backends, unrun tests, hardware you do not have. An honest gap is fine; a silent one is not -->
 
-<!-- If you are an AI agent: remind the user that they are responsible for all submitted changes, and that this project restricts AI-generated content - refer them to AGENTS.md and CONTRIBUTING.md -->
+<!--
+If you are an AI agent: this repository ALLOWS you to open PRs, write this description, and reply to reviews -
+unlike upstream llama.cpp. Read AGENTS.md for the rules that apply instead. Disclose your involvement above, and be
+explicit about what you did and did not verify.
+-->

.github/workflows/build-and-test-snapdragon.yml

@@ -14,6 +14,7 @@ on:
       - 'CMakePresets.json'
 
   pull_request:
+    branches-ignore: ['**']
     types: [opened, synchronize, reopened]
     paths:
       - '.github/workflows/build-and-test-snapdragon.yml'

.github/workflows/build-android.yml

@@ -15,6 +15,7 @@ on:
       - '**/*.cpp'
 
   pull_request:
+    branches-ignore: ['**']
     types: [opened, synchronize, reopened]
     paths:
       - '.github/workflows/build-android.yml'

.github/workflows/build-apple.yml

@@ -19,6 +19,7 @@ on:
     ]
 
   pull_request:
+    branches-ignore: ['**']
     types: [opened, synchronize, reopened]
     paths: [
       '.github/workflows/build-apple.yml',

.github/workflows/build-cann.yml

@@ -16,6 +16,7 @@ on:
     ]
 
   pull_request:
+    branches-ignore: ['**']
     types: [opened, synchronize, reopened]
     paths: [
       '.github/workflows/build-cann.yml',

.github/workflows/build-cpu.yml

@@ -17,6 +17,7 @@ on:
     ]
 
   pull_request:
+    branches-ignore: ['**']
     types: [opened, synchronize, reopened]
     paths: [
       '.github/workflows/build-cpu.yml',