Fix the failing Windows build

.github/workflows/release.yml

@@ -75,27 +75,27 @@ jobs:
         include:
           - os: ubuntu-latest
             target: x86_64-linux-gnu
-            artifact_name: zigformer-cli-linux-x86_64.zip
+            artifact_name: zigformer-linux-x86_64.zip
             asset_content_type: application/zip
           - os: ubuntu-latest
             target: aarch64-linux-gnu
-            artifact_name: zigformer-cli-linux-aarch64.zip
+            artifact_name: zigformer-linux-aarch64.zip
             asset_content_type: application/zip
           - os: macos-latest
             target: aarch64-macos # Apple Silicon (ARM64)
-            artifact_name: zigformer-cli-macos-aarch64.zip
+            artifact_name: zigformer-macos-aarch64.zip
             asset_content_type: application/zip
           - os: macos-latest
             target: x86_64-macos # Intel
-            artifact_name: zigformer-cli-macos-x86_64.zip
+            artifact_name: zigformer-macos-x86_64.zip
             asset_content_type: application/zip
           - os: windows-latest
             target: x86_64-windows-msvc
-            artifact_name: zigformer-cli-windows-x86_64.zip
+            artifact_name: zigformer-windows-x86_64.zip
             asset_content_type: application/zip
           - os: windows-latest
             target: aarch64-windows-msvc
-            artifact_name: zigformer-cli-windows-aarch64.zip
+            artifact_name: zigformer-windows-aarch64.zip
             asset_content_type: application/zip
 
     steps:
@@ -120,6 +120,11 @@ jobs:
             ls -l zig-out/bin || true
             exit 1
           fi
+          if [ ! -f zig-out/bin/zigformer-gui ]; then
+            echo "ERROR: built binary zig-out/bin/zigformer-gui not found"
+            ls -l zig-out/bin || true
+            exit 1
+          fi
 
       - name: Verify built binary exists (Windows)
         if: runner.os == 'Windows'
@@ -129,13 +134,19 @@ jobs:
             dir zig-out\\bin
             exit 1
           }
+          if (!(Test-Path "zig-out\\bin\\zigformer-gui.exe")) {
+            Write-Host 'ERROR: built binary zig-out\\bin\\zigformer-gui.exe not found'
+            dir zig-out\\bin
+            exit 1
+          }
 
       - name: Prepare artifact for Linux/macOS
         if: runner.os != 'Windows'
         run: |
           RELEASE_DIR="release-${{ matrix.target }}-${{ github.run_id }}"
           mkdir -p "$RELEASE_DIR"
           mv zig-out/bin/zigformer-cli "$RELEASE_DIR"/
+          mv zig-out/bin/zigformer-gui "$RELEASE_DIR"/
           # Create a zip of the release contents (avoid nesting an extra top-level directory)
           (cd "$RELEASE_DIR" && zip -r "../${{ matrix.artifact_name }}" .)
 
@@ -145,6 +156,7 @@ jobs:
           $env:RELEASE_DIR = "release-${{ matrix.target }}-${{ github.run_id }}"
           New-Item -ItemType Directory -Path $env:RELEASE_DIR -Force | Out-Null
           Move-Item zig-out\bin\zigformer-cli.exe $env:RELEASE_DIR\
+          Move-Item zig-out\bin\zigformer-gui.exe $env:RELEASE_DIR\
           Compress-Archive -Path "$env:RELEASE_DIR\*" -DestinationPath ${{ matrix.artifact_name }}
 
       - name: Upload Release Asset

README.md

@@ -28,7 +28,7 @@ and [nanochat](https://github.com/karpathy/nanochat) projects, and follows the a
 ["Language Models are Unsupervised Multitask Learners"](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)
 papers.
 It can be used as a [Zig library](https://CogitatorTech.github.io/zigformer/) for building LLMs or as a
-[standalone application](https://github.com/CogitatorTech/zigformer/releases)
+[standalone application](https://github.com/CogitatorTech/zigformer/releases/latest)
 for training, inference, and chatting with the model.
 
 The diagrams below show the high-level architecture and its core components.
@@ -171,7 +171,7 @@ zig build run -- predict --prompt "How do mountains form?" --top-k 5 --load-mode
 zig build run-gui -- --load-model model.bin
 ```
 
-![ZigFormer Web UI](assets/screenshots/zigformer_webui_v0.1.0.jpeg)
+![ZigFormer Web UI](assets/screenshots/zigformer_webui_v0.1.1.jpeg)
 
 ---
 

build.zig.zon

@@ -1,6 +1,6 @@
 .{
     .name = .zigformer,
-    .version = "0.1.0",
+    .version = "0.1.1",
     .fingerprint = 0xe046275379ddc4d8, // Changing this has security and trust implications.
     .minimum_zig_version = "0.15.2",
     .dependencies = .{

src/cli.zig

@@ -216,7 +216,7 @@ fn trainAndMaybeRepl(allocator: std.mem.Allocator, pretrain_path: []const u8, ch
     std.debug.print("Type a prompt and press Enter to generate text.\n", .{});
     std.debug.print("Type 'exit' to quit.\n", .{});
 
-    const stdin_file = std.fs.File{ .handle = std.posix.STDIN_FILENO };
+    const stdin_file = std.fs.File.stdin();
     const stdin = stdin_file.deprecatedReader();
     var buffer: [1024]u8 = undefined;
     while (true) {
@@ -245,33 +245,6 @@ fn execRoot(ctx: chilli.CommandContext) !void {
     var config = if (config_path.len > 0) try loadConfig(allocator, config_path) else Config{};
     defer config.deinit(allocator);
 
-    // Override with CLI flags if provided (checking if they are different from defaults is tricky with chilli,
-    // so we'll assume CLI flags take precedence if they are set to non-default values or if we just use them directly.
-    // Actually, a better approach is: use config as base, then overwrite with CLI flags.
-    // But chilli returns defaults if flag is missing.
-    // So we need to know if flag was actually passed. Chilli doesn't easily expose this.
-    // For now, let's just use CLI flags if config is NOT present, OR if we want to support overrides,
-    // we have to accept that CLI defaults might overwrite config values.
-    // To solve this properly:
-    // 1. Load config
-    // 2. For each field, check if CLI flag was passed (not easy with current chilli usage).
-    // Alternative: We only use config if --config is passed, and ignore other flags? No, overrides are good.
-    // Let's assume: Config file sets defaults. CLI flags override.
-    // But chilli returns default values if flag is missing.
-    // So if config has batch_size=64, and CLI default is 32, and user runs without --batch-size, chilli returns 32.
-    // If we overwrite config with 32, we lose the config value.
-    // We need to check if the flag was present.
-    // Since we can't easily do that, let's prioritize CLI flags ONLY if they are explicitly different from our hardcoded defaults?
-    // Or simpler: If --config is passed, we use it. We can manually parse args to see if flags are present, but that's messy.
-
-    // Let's stick to the plan: Config file sets values. CLI flags override.
-    // If we want CLI to override, we need to know if user typed it.
-    // Given the constraints, let's do this:
-    // If --config is present, use it.
-    // AND we will NOT read other flags if --config is present, to avoid confusion.
-    // OR we can say: CLI flags are ignored if --config is present, EXCEPT for interactive/save-model maybe?
-    // Let's go with: If --config is present, it is the source of truth.
-
     if (config_path.len > 0) {
         std.debug.print("Loaded configuration from {s}\n", .{config_path});
     } else {
@@ -361,7 +334,7 @@ pub fn main() anyerror!void {
     var root_cmd = try chilli.Command.init(allocator, .{
         .name = "zigformer-cli",
         .description = "An educational transformer-based LLM in Zig",
-        .version = "v0.1.0",
+        .version = "v0.1.1",
         .exec = execRoot,
     });
     defer root_cmd.deinit();

src/gui.zig

@@ -27,7 +27,7 @@ const Config = struct {
     }
 };
 
-// Global state for the server
+// Global state holders for the server
 const ServerState = struct {
     allocator: std.mem.Allocator,
     model: *llm.LLM,
@@ -421,7 +421,7 @@ pub fn main() !void {
     var cmd = try chilli.Command.init(allocator, .{
         .name = "zigformer-gui",
         .description = "Web GUI for ZigFormer",
-        .version = "0.1.0",
+        .version = "0.1.1",
         .exec = execGui,
     });
     defer cmd.deinit();

src/gui/index.html

@@ -381,7 +381,7 @@
 
             <div class="control">
                 <label>
-                    Top-k
+                    Top-p
                     <span class="value-display" id="top-p-val">0.0</span>
                 </label>
                 <input id="top-p" max="1" min="0" step="0.05" type="range" value="0">

src/lib.zig

@@ -1,6 +1,6 @@
 //! ## ZigFormer
 //!
-//! ZigFormer is an implementation of a transformer-based language model (like GPT-2)
+//! ZigFormer is an implementation of a transformer-based language model (LLM)
 //! written in pure Zig. It provides a framework for understanding and experimenting with
 //! transformer architectures.
 //!
@@ -40,7 +40,7 @@
 //! - `self_attention`: Multi-head self-attention mechanism
 //! - `feed_forward`: Position-wise feed-forward networks
 //! - `layer_norm`: Layer normalization
-//! - `transformer`: Transformer block (attention and FFN)
+//! - `transformer`: Transformer block (feed-forward and attention layers)
 //! - `output_projection`: Output layer for vocabulary prediction
 //! - `llm`: Language model implementation with training and inference API
 //! - `optimizer`: Adam optimizer

src/zigformer/llm.zig

@@ -674,24 +674,6 @@ pub const LLM = struct {
     }
 
     pub fn setBatchSize(self: *LLM, batch_size: usize) void {
-        // Iterate over layers and set batch_size for SelfAttention layers
-        // Note: We need to know which layers are SelfAttention.
-        // In our simple structure, we know layers 1, 2, 3 are TransformerBlocks.
-        // TransformerBlock contains SelfAttention.
-        // But Layer is type-erased.
-        // Ideally, we should add setBatchSize to Layer vtable, but that's a big change.
-        // For now, we'll rely on the known structure and pointer casting, which is risky but fits the current style.
-        // Actually, TransformerBlock has a setBatchSize method we should add.
-        // Let's assume we add setBatchSize to TransformerBlock and call it here.
-
-        // Wait, we can't easily cast opaque pointers back to types without RTTI or knowing the type.
-        // Given the fixed structure:
-        // Layer 0: Embeddings (no batch_size needed)
-        // Layer 1: TransformerBlock
-        // Layer 2: TransformerBlock
-        // Layer 3: TransformerBlock
-        // Layer 4: OutputProjection (no batch_size needed)
-
         const embeddings: *Embeddings = @ptrCast(@alignCast(self.network.items[0].self));
         embeddings.setBatchSize(batch_size);
 

src/zigformer/self_attention.zig

@@ -635,13 +635,6 @@ test "SelfAttention Causal Masking with Cache" {
     var output = try attn.forward(input, true);
     defer output.deinit();
 
-    // Check cached_attention_scores
-    // Should be (num_heads * seq_len) x seq_len
-    // For each head, the 4x4 score matrix should be lower triangular (masked).
-    // scores[i, j] should be -inf (or very small after softmax, but we check pre-softmax scores if possible?
-    // Wait, cached_attention_scores stores POST-softmax scores.
-    // So masked values should be 0.0.
-
     const scores = attn.cached_attention_scores;
     // Rows: num_heads * seq_len
     // Cols: seq_len (since cache_len = seq_len after first pass)