examples/multimodal/index.html (1)

244-244: ⚡ Quick win

Avoid logging every streamed chunk in the hot path.

Line [244] logs each chunk during generation; this can reduce measured token throughput and adds noisy console output. Consider gating behind a debug flag or removing it.

Suggested fix

+      const DEBUG_STREAM_CHUNKS = false;
...
-            console.log('Received chunk:', chunk);
+            if (DEBUG_STREAM_CHUNKS) console.log('Received chunk:', chunk);

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@examples/multimodal/index.html` at line 244, The hot-path
console.log('Received chunk:', chunk) inside the streaming callback is too noisy
and harms throughput; remove that unconditional log or gate it behind a runtime
debug flag (e.g., DEBUG or verbose) and only log when the flag is true so
streaming performance isn't impacted—locate the streaming callback that receives
the chunk variable in examples/multimodal/index.html and replace the direct
console.log with a conditional log controlled by the debug flag.

scripts/docker-compose.yml (1)

24-34: 💤 Low value

Consider cleanup and version documentation.

The Dawn package integration looks correct. A couple of minor suggestions:

1. The emdawn.zip file remains after extraction. Consider adding rm emdawn.zip after line 32 to save disk space during builds.
2. Consider adding a comment explaining how to update DAWN_TAG when newer versions are needed.

♻️ Optional cleanup

         python3 -c "import zipfile; zf=zipfile.ZipFile('emdawn.zip','r'); zf.extractall('/source/build/emdawn'); zf.close()"
+        rm emdawn.zip

         emcmake cmake .. -DGGML_WEBGPU=ON -DGGML_WEBGPU_JSPI=ON -DEMDAWNWEBGPU_DIR="$${EMDAWNWEBGPU_DIR}"

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@scripts/docker-compose.yml` around lines 24 - 34, Add cleanup and a brief
version-update note: after extracting emdawn.zip (the python3 zipfile.extractall
call that creates emdawn.zip), remove the downloaded archive (rm emdawn.zip) to
free build disk space, and add a comment above the DAWN_TAG declaration
explaining how to update DAWN_TAG/EMDAWN_PKG (e.g., where to find new release
tags or naming pattern) so future maintainers know how to bump the Dawn version
and regenerate EMDAWN_PKG/EMDAWNWEBGPU_DIR.

vitest.config.ts (1)

29-47: 💤 Low value

Consider documenting that WebGPU tests require Chromium.

The providerOptions logic correctly handles WebGPU mode with Chromium-specific flags, but when WEBGPU=1 is set alongside BROWSER=safari, the config will use Chromium launch args with the playwright provider anyway (since the WEBGPU check comes first). This is likely intentional since Safari doesn't support the enable-unsafe-webgpu flag, but a brief comment would clarify the expected behavior.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@vitest.config.ts` around lines 29 - 47, Add a short inline comment above the
providerOptions conditional explaining that when WEBGPU is true the config
intentionally uses Chromium-specific launch args (chromeArgsWebGPU) and will
override BROWSER=safari/SAFARI because Safari does not support the
enable-unsafe-webgpu flag; reference WEBGPU, SAFARI, and chromeArgsWebGPU so
future readers understand this precedence and why Chromium is required for
WebGPU tests.

src/wllama.ts (1)

473-474: 💤 Low value

Document the default behavior change for n_gpu_layers.

The default value of n_gpu_layers: 99999 means WebGPU offloading is enabled by default when available. This is a significant behavior change that users should be aware of. Consider adding a note in the LoadModelParams type documentation or README to clarify that GPU offloading is now opt-out rather than opt-in.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/wllama.ts` around lines 473 - 474, Update the documentation to call out
that n_gpu_layers now defaults to 99999 (in the LoadModelParams type and any
public API docs/README), explaining this enables WebGPU/GPU offloading by
default when available (i.e., it is opt-out not opt-in); add a short JSDoc
comment on the LoadModelParams type and/or the function that consumes it (where
n_gpu_layers is used) noting the default and how to disable offloading (set
n_gpu_layers to 0 or an explicit value), and update any README/usage examples to
reflect the new default behavior.

src/workers-code/llama-cpp.js (1)

247-251: 💤 Low value

Minor: Consider adding explicit await for zero-arg async functions.

For consistency, the zero-arg branch (line 250) could explicitly await when isAsync is true, matching the two-arg branch pattern. While the current code works correctly due to async function semantics, explicit awaiting makes the intent clearer:

       if (args.length === 2) {
         result = isAsync ? await fn(action, req) : fn(action, req);
       } else {
-        result = fn();
+        result = isAsync ? await fn() : fn();
       }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/workers-code/llama-cpp.js` around lines 247 - 251, The zero-arg call path
currently does not explicitly await async functions; update the args.length ===
2 else branch so that when isAsync is true you await fn() (e.g., replace result
= fn() with result = isAsync ? await fn() : fn()), keeping the existing
variables fn, isAsync and result and preserving the surrounding async context
used where this args.length check occurs.

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes