-
Use greedy decoding: do_sample sampling path fails on fp16 logits in transformers.js 3.8.1; remove temp instrumentation
-
Instrument adapter (temporary): expose generator + wrapper call counter
-
Adapter v3: learn expected KV dtypes via try/flip/retry (session metadata unavailable in CDN ORT build)
-
Add KV dtype adapter: align KV-cache feed dtypes with session input metadata (handles both fp16- and fp32-resolving ORT builds)
-
Force WASM inference: WebGPU shader-f16 corrupts this fp16 export (verified on-device); WASM output verified correct. Trim max_new_tokens to 160.
-
Fix generation: remove fp16 I/O shim (config kv_cache_dtype handles it), pin transformers.js 3.8.1, refresh stale cached configs, restore WASM fallback with shader-f16 gate
-
Shim: only cast when session input metadata expects float16
-
Add fp16 I/O shim: cast fp32 KV feeds to fp16 and fp16 logits to fp32
-
Use current Transformers.js 3.x for fp16 KV cache
-
Require WebGPU for fp16 GPU Mart inference
-
Use fp16 ONNX with graph optimizations disabled
-
Convert to free static Space: browser-side model + voice/text UI