Apply the GGUF's suppress tokens in the WebGPU TTS sampler

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 ## Unreleased
 
+- WebGPU text-to-speech applies the GGUF's suppress tokens in the talker sampler, as upstream does.
+  Qwen3-TTS no longer fails with `code out of range` when a codec control token is sampled.
 - `createCompletion` can explicitly count prompt tokens in penalties with a grammar using `penaltyPromptTokensWithGrammar`; a matching loaded-core capability is required, and existing grammar defaults stay unchanged.
 
 - Large wasm32 streamed model files use bounded size hints to reserve filesystem capacity, avoiding geometric file-growth allocation failures at 1 GiB without treating hidden compression metadata as a decoded-size contract.

src/llama_webgpu_tts.cpp

@@ -109,12 +109,32 @@ uint32_t capabilities_for(
            : 0u);
 }
 
-llama_sampler * init_sampler(const llama_webgpu_tts_request & request) {
+llama_sampler * init_sampler(
+    const llama_vocab * vocab,
+    const llama_webgpu_tts_request & request) {
   llama_sampler * sampler =
       llama_sampler_chain_init(llama_sampler_chain_default_params());
   if (sampler == nullptr) {
     return nullptr;
   }
+  // As in upstream common_sampler, suppressed tokens (for Qwen3-TTS the codec's
+  // control tokens, which code2wav rejects as out-of-range codes) get a -inf bias first.
+  int32_t n_suppress = 0;
+  const llama_token * suppress =
+      llama_vocab_get_suppress_tokens(vocab, &n_suppress);
+  if (suppress != nullptr && n_suppress > 0) {
+    std::vector<llama_logit_bias> bias;
+    bias.reserve(static_cast<size_t>(n_suppress));
+    for (int32_t i = 0; i < n_suppress; ++i) {
+      bias.push_back({suppress[i], -INFINITY});
+    }
+    llama_sampler_chain_add(
+        sampler,
+        llama_sampler_init_logit_bias(
+            llama_vocab_n_tokens(vocab),
+            static_cast<int32_t>(bias.size()),
+            bias.data()));
+  }
   llama_sampler_chain_add(sampler, llama_sampler_init_top_k(request.top_k));
   llama_sampler_chain_add(sampler, llama_sampler_init_top_p(request.top_p, 1));
   llama_sampler_chain_add(sampler, llama_sampler_init_min_p(request.min_p, 1));
@@ -307,7 +327,8 @@ llama_webgpu_tts_status llama_webgpu_tts_start(
     tts->speaker = wrapper.bitmap;
   }
 
-  tts->sampler = init_sampler(request);
+  tts->sampler = init_sampler(
+      llama_model_get_vocab(llama_get_model(tts->llama)), request);
   if (tts->sampler == nullptr) {
     return fail(
         tts,

tests/bridge/tts_sampler_suppress_tokens_test.mjs

@@ -0,0 +1,36 @@
+// The TTS talker sampler must bias the GGUF's suppress tokens to -inf before any
+// other stage, as upstream common_sampler does. Qwen3-TTS lists its codec control
+// tokens there, and code2wav rejects a sampled control token as out of range.
+import assert from 'node:assert/strict';
+import test from 'node:test';
+
+import { readRepoText } from './bridge_js_source.mjs';
+
+const TTS = readRepoText('src/llama_webgpu_tts.cpp');
+
+function samplerBody() {
+  const start = TTS.indexOf('llama_sampler * init_sampler(');
+  assert.notEqual(start, -1, 'init_sampler not found');
+  const end = TTS.indexOf('\nllama_webgpu_tts_status ', start);
+  assert.ok(end > start, 'init_sampler end not found');
+  return TTS.slice(start, end);
+}
+
+test('TTS sampler biases suppress tokens to -inf before the first sampling stage', () => {
+  const body = samplerBody();
+  const suppress = body.indexOf('llama_vocab_get_suppress_tokens(');
+  const bias = body.indexOf('llama_sampler_init_logit_bias(');
+  const firstStage = body.indexOf('llama_sampler_init_top_k(');
+  assert.ok(suppress !== -1, 'suppress tokens are not read from the vocab');
+  assert.ok(bias > suppress, 'logit bias is not built from the suppress tokens');
+  assert.ok(firstStage > bias, 'logit bias must be added before top_k');
+  assert.match(body, /-INFINITY/);
+  assert.match(body, /llama_vocab_n_tokens\(vocab\)/);
+});
+
+test('TTS sampler is built from the loaded model vocab', () => {
+  assert.match(
+    TTS,
+    /init_sampler\(\s*llama_model_get_vocab\(llama_get_model\(tts->llama\)\),\s*request\)/,
+  );
+});