@sauravpanda/flare 0.1.0

package/pkg/flare_web.js

@@ -0,0 +1,2790 @@
+/* @ts-self-types="./flare_web.d.ts" */
+
+/**
+ * Flare LLM inference engine, exported to JS.
+ *
+ * Holds a loaded model and runs greedy/sampled token generation.
+ * The detected chat template is available via `chat_template_name` and
+ * `apply_chat_template` so the browser demo can format prompts correctly
+ * for instruction-tuned models.
+ */
+export class FlareEngine {
+    static __wrap(ptr) {
+        ptr = ptr >>> 0;
+        const obj = Object.create(FlareEngine.prototype);
+        obj.__wbg_ptr = ptr;
+        FlareEngineFinalization.register(obj, obj.__wbg_ptr, obj);
+        return obj;
+    }
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        FlareEngineFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_flareengine_free(ptr, 0);
+    }
+    /**
+     * Whether the model requests automatic BOS token prepending.
+     *
+     * Sourced from `tokenizer.ggml.add_bos_token` in the GGUF metadata.
+     * When `true`, all generation methods (`generate_tokens`, `begin_stream`,
+     * `generate_text`, `generate_stream`) automatically prepend the BOS token
+     * to the input token sequence unless it is already the first token.
+     * @returns {boolean}
+     */
+    get add_bos_token() {
+        const ret = wasm.flareengine_add_bos_token(this.__wbg_ptr);
+        return ret !== 0;
+    }
+    /**
+     * Register a stop sequence.
+     *
+     * Generation halts (without emitting the matched tokens) as soon as the
+     * decoded output ends with `sequence`.  Call once per stop string before
+     * `begin_stream` or `generate_with_params`.
+     *
+     * Stop sequences are cleared by `reset()` or `clear_stop_sequences()`.
+     *
+     * ```javascript
+     * engine.add_stop_sequence("<|im_end|>");
+     * engine.add_stop_sequence("</s>");
+     * engine.begin_stream_with_params(promptIds, 200, 0.8, 0.95, 40, 1.1);
+     * ```
+     * @param {string} sequence
+     */
+    add_stop_sequence(sequence) {
+        const ptr0 = passStringToWasm0(sequence, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        wasm.flareengine_add_stop_sequence(this.__wbg_ptr, ptr0, len0);
+    }
+    /**
+     * Format a user message (and optional system prompt) using the model's
+     * auto-detected chat template.  Returns the formatted prompt string ready
+     * to be passed to `FlareTokenizer.encode()`.
+     *
+     * Pass an empty string for `system_message` to omit the system turn.
+     *
+     * # JS example
+     * ```javascript
+     * const prompt = engine.apply_chat_template(
+     *   'Explain quantum computing in simple terms.',
+     *   'You are a helpful assistant.'
+     * );
+     * const ids = tokenizer.encode(prompt);
+     * const output = engine.generate_tokens(ids, 128);
+     * ```
+     * @param {string} user_message
+     * @param {string} system_message
+     * @returns {string}
+     */
+    apply_chat_template(user_message, system_message) {
+        let deferred3_0;
+        let deferred3_1;
+        try {
+            const ptr0 = passStringToWasm0(user_message, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len0 = WASM_VECTOR_LEN;
+            const ptr1 = passStringToWasm0(system_message, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len1 = WASM_VECTOR_LEN;
+            const ret = wasm.flareengine_apply_chat_template(this.__wbg_ptr, ptr0, len0, ptr1, len1);
+            deferred3_0 = ret[0];
+            deferred3_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred3_0, deferred3_1, 1);
+        }
+    }
+    /**
+     * Model architecture name from `general.architecture` in the GGUF metadata.
+     *
+     * Returns a lowercase string such as `"llama"`, `"mistral"`, `"gemma2"`,
+     * `"phi3"`, or `"qwen2"`. Returns `"unknown"` if the field is absent.
+     * @returns {string}
+     */
+    get architecture() {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_architecture(this.__wbg_ptr);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Prepare for token-by-token streaming.
+     *
+     * Runs the prefill pass on `prompt_tokens`, then initialises internal
+     * state so that subsequent calls to `next_token()` each produce one
+     * output token.  Call `engine.reset()` before `begin_stream()` to start
+     * a fresh conversation.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * engine.begin_stream(promptIds, 128);
+     * function tick() {
+     *   const id = engine.next_token();
+     *   if (id === undefined) { /* done */ return; }
+     *   output.textContent += tokenizer.decode_one(id);
+     *   requestAnimationFrame(tick);   // yield to browser, then continue
+     * }
+     * requestAnimationFrame(tick);
+     * ```
+     * @param {Uint32Array} prompt_tokens
+     * @param {number} max_tokens
+     */
+    begin_stream(prompt_tokens, max_tokens) {
+        const ptr0 = passArray32ToWasm0(prompt_tokens, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        wasm.flareengine_begin_stream(this.__wbg_ptr, ptr0, len0, max_tokens);
+    }
+    /**
+     * Begin a token-by-token stream, healing the last prompt token.
+     *
+     * Identical to `begin_stream` but avoids double-processing the final prompt
+     * token: the prefill runs only tokens `[0 .. n-2]`, then the first
+     * `next_token()` call processes the last prompt token at its correct
+     * position `n-1` and produces the first output token.  This keeps RoPE
+     * positional embeddings consistent and is recommended when the prompt
+     * ends at a natural token boundary (e.g. when encoding a user turn in a
+     * chat template).
+     *
+     * Falls back to `begin_stream` for prompts shorter than 2 tokens.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * const ids = engine.encode_text(engine.apply_chat_template(userMsg, sysMsg));
+     * engine.begin_stream_healed(ids, 256);
+     * requestAnimationFrame(function tick() {
+     *   const id = engine.next_token();
+     *   if (id !== undefined) output.textContent += tokenizer.decode_one(id);
+     *   if (!engine.stream_done) requestAnimationFrame(tick);
+     * });
+     * ```
+     * @param {Uint32Array} prompt_tokens
+     * @param {number} max_tokens
+     */
+    begin_stream_healed(prompt_tokens, max_tokens) {
+        const ptr0 = passArray32ToWasm0(prompt_tokens, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        wasm.flareengine_begin_stream_healed(this.__wbg_ptr, ptr0, len0, max_tokens);
+    }
+    /**
+     * Like `begin_stream_healed` but with full sampling parameters.
+     *
+     * Combines position-consistent prefill (see `begin_stream_healed`) with
+     * the same temperature / top-p / top-k / repeat-penalty / min-p controls
+     * available in `begin_stream_with_params`.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * const ids = engine.encode_text(engine.apply_chat_template(userMsg, sysMsg));
+     * engine.begin_stream_healed_with_params(ids, 256, 0.8, 0.95, 40, 1.1, 0.0);
+     * requestAnimationFrame(function tick() {
+     *   const id = engine.next_token();
+     *   if (id !== undefined) output.textContent += tokenizer.decode_one(id);
+     *   if (!engine.stream_done) requestAnimationFrame(tick);
+     * });
+     * ```
+     * @param {Uint32Array} prompt_tokens
+     * @param {number} max_tokens
+     * @param {number} temperature
+     * @param {number} top_p
+     * @param {number} top_k
+     * @param {number} repeat_penalty
+     * @param {number} min_p
+     */
+    begin_stream_healed_with_params(prompt_tokens, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p) {
+        const ptr0 = passArray32ToWasm0(prompt_tokens, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        wasm.flareengine_begin_stream_healed_with_params(this.__wbg_ptr, ptr0, len0, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p);
+    }
+    /**
+     * Like `begin_stream` but with temperature / top-p sampling.
+     *
+     * `temperature`: 0.0 = greedy, 0.7–1.0 = typical creative range.
+     * `top_p`: nucleus sampling threshold (0.0–1.0); 0.9 is a good default.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * engine.begin_stream_with_params(promptIds, 128, 0.8, 0.9);
+     * function tick() {
+     *   const id = engine.next_token();
+     *   if (id === undefined) return;
+     *   output.textContent += tokenizer.decode_one(id);
+     *   requestAnimationFrame(tick);
+     * }
+     * requestAnimationFrame(tick);
+     * ```
+     * Begin a token-by-token stream with sampling parameters including top-k.
+     *
+     * - `temperature`: controls randomness (0 = greedy, higher = more random)
+     * - `top_p`: nucleus sampling — keep the smallest token set whose cumulative
+     *   probability ≥ `top_p` (1.0 = disabled; applied when < 1.0)
+     * - `top_k`: keep only the `top_k` highest-probability tokens before sampling
+     *   (0 = disabled; applied when `top_p` is 1.0 and `top_k` > 0)
+     * - `repeat_penalty`: penalty applied to logits of recently-seen tokens to
+     *   reduce repetition (1.0 = disabled, 1.1–1.3 = typical range)
+     *
+     * ```javascript
+     * engine.begin_stream_with_params(promptIds, 200, 0.8, 0.95, 40, 1.1, 0.0);
+     * ```
+     * @param {Uint32Array} prompt_tokens
+     * @param {number} max_tokens
+     * @param {number} temperature
+     * @param {number} top_p
+     * @param {number} top_k
+     * @param {number} repeat_penalty
+     * @param {number} min_p
+     */
+    begin_stream_with_params(prompt_tokens, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p) {
+        const ptr0 = passArray32ToWasm0(prompt_tokens, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        wasm.flareengine_begin_stream_with_params(this.__wbg_ptr, ptr0, len0, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p);
+    }
+    /**
+     * BOS (beginning of sequence) token ID from the GGUF model metadata, if present.
+     * Some models require this to be prepended to the input token sequence.
+     * @returns {number | undefined}
+     */
+    get bos_token_id() {
+        const ret = wasm.flareengine_bos_token_id(this.__wbg_ptr);
+        return ret === 0x100000001 ? undefined : ret;
+    }
+    /**
+     * Name of the auto-detected chat template (e.g. `"ChatML"`, `"Llama3"`,
+     * `"Alpaca"`, `"Raw"`).  Use this to display the template in the UI and
+     * decide whether to call `apply_chat_template` before encoding.
+     * @returns {string}
+     */
+    get chat_template_name() {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_chat_template_name(this.__wbg_ptr);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Clear any previously loaded raw quantized weights.
+     *
+     * After calling this the engine uses the f32 dequantized path for all
+     * matrix operations until `load_raw_weights` is called again.
+     */
+    clear_raw_weights() {
+        wasm.flareengine_clear_raw_weights(this.__wbg_ptr);
+    }
+    /**
+     * Remove all registered stop sequences.
+     */
+    clear_stop_sequences() {
+        wasm.flareengine_clear_stop_sequences(this.__wbg_ptr);
+    }
+    /**
+     * Compute the perplexity of `text` under the loaded model.
+     *
+     * Encodes `text` with the embedded GGUF vocabulary, runs one forward pass
+     * per token, and measures the log-probability of each correct next-token
+     * prediction.  Perplexity = exp(−mean(log_probs)).
+     *
+     * The KV cache is reset **before and after** the evaluation so the engine
+     * returns to a clean state.
+     *
+     * Returns `f32::INFINITY` if the text encodes to fewer than 2 tokens or if
+     * no GGUF vocabulary is available.
+     *
+     * # JS example
+     * ```javascript
+     * const ppl = engine.compute_perplexity("The quick brown fox");
+     * console.log("Perplexity:", ppl);
+     * ```
+     * @param {string} text
+     * @returns {number}
+     */
+    compute_perplexity(text) {
+        const ptr0 = passStringToWasm0(text, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_compute_perplexity(this.__wbg_ptr, ptr0, len0);
+        return ret;
+    }
+    /**
+     * Fraction of the context window consumed (0.0 = empty, 1.0 = full).
+     *
+     * Equivalent to `tokens_used / max_seq_len`. Returns 0.0 if `max_seq_len` is 0.
+     * @returns {number}
+     */
+    get context_window_pct() {
+        const ret = wasm.flareengine_context_window_pct(this.__wbg_ptr);
+        return ret;
+    }
+    /**
+     * Count the number of tokens in `text` using the model's embedded GGUF vocabulary.
+     *
+     * Returns 0 if the model was not loaded from a GGUF file (e.g. SafeTensors only).
+     *
+     * # JS example
+     * ```javascript
+     * const n = engine.count_tokens(textarea.value);
+     * counter.textContent = `${n} / ${engine.max_seq_len} tokens`;
+     * ```
+     * @param {string} text
+     * @returns {number}
+     */
+    count_tokens(text) {
+        const ptr0 = passStringToWasm0(text, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_count_tokens(this.__wbg_ptr, ptr0, len0);
+        return ret >>> 0;
+    }
+    /**
+     * Decode token IDs to text using the embedded GGUF vocabulary.
+     *
+     * Returns an empty string if no GGUF vocab is available.
+     *
+     * # JS example
+     * ```javascript
+     * const text = engine.decode_ids(generatedIds);
+     * ```
+     * @param {Uint32Array} ids
+     * @returns {string}
+     */
+    decode_ids(ids) {
+        let deferred2_0;
+        let deferred2_1;
+        try {
+            const ptr0 = passArray32ToWasm0(ids, wasm.__wbindgen_malloc);
+            const len0 = WASM_VECTOR_LEN;
+            const ret = wasm.flareengine_decode_ids(this.__wbg_ptr, ptr0, len0);
+            deferred2_0 = ret[0];
+            deferred2_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred2_0, deferred2_1, 1);
+        }
+    }
+    /**
+     * Decode a single token ID to its text piece.
+     *
+     * Convenience wrapper around `decode_ids` for use directly inside a
+     * `next_token()` loop so callers don't need a separate `FlareTokenizer`.
+     *
+     * Returns an empty string if no GGUF vocab is loaded.
+     *
+     * # JS example
+     * ```javascript
+     * engine.begin_stream(promptIds, 128);
+     * requestAnimationFrame(function tick() {
+     *   const id = engine.next_token();
+     *   if (id !== undefined) output.textContent += engine.decode_token(id);
+     *   if (!engine.stream_done) requestAnimationFrame(tick);
+     * });
+     * ```
+     * @param {number} id
+     * @returns {string}
+     */
+    decode_token(id) {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_decode_token(this.__wbg_ptr, id);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Decode a single token ID, correctly handling multi-byte UTF-8 sequences.
+     *
+     * SentencePiece tokenizers encode non-ASCII characters as consecutive
+     * byte-level tokens such as `<0xE4>`, `<0xB8>`, `<0xAD>` (the UTF-8
+     * encoding of `中`).  The basic `decode_token` function returns incorrect
+     * Latin-1 characters in these cases because it treats each byte as an
+     * independent Unicode scalar.
+     *
+     * `decode_token_chunk` accumulates bytes in an internal buffer until a
+     * complete, valid UTF-8 sequence is assembled, then returns it as a
+     * `String`.  While the sequence is incomplete it returns an empty string,
+     * and when a regular (non-byte) token is encountered it flushes any
+     * buffered bytes (replacing invalid sequences with U+FFFD) before
+     * returning the decoded text.
+     *
+     * **Use this instead of `decode_token` whenever you are streaming tokens
+     * that may include non-Latin characters.**
+     *
+     * ```javascript
+     * engine.begin_stream(prompt, 256);
+     * function tick() {
+     *   const id = engine.next_token();
+     *   if (id !== undefined) output.textContent += engine.decode_token_chunk(id);
+     *   if (!engine.stream_done) requestAnimationFrame(tick);
+     * }
+     * requestAnimationFrame(tick);
+     * ```
+     * @param {number} id
+     * @returns {string}
+     */
+    decode_token_chunk(id) {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_decode_token_chunk(this.__wbg_ptr, id);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Look up the token embedding row for `token_id` as a flat `Float32Array`.
+     *
+     * The length of the returned vector is `hidden_dim`. See also
+     * [`FlareEngine::output_projection`] for the inverse tail step.
+     * @param {number} token_id
+     * @returns {Float32Array}
+     */
+    embed_token(token_id) {
+        const ret = wasm.flareengine_embed_token(this.__wbg_ptr, token_id);
+        var v1 = getArrayF32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v1;
+    }
+    /**
+     * # JS example
+     * ```javascript
+     * const ids = engine.encode_text("Hello, world!");
+     * const output = engine.generate_tokens(ids, 64);
+     * ```
+     * @param {string} text
+     * @returns {Uint32Array}
+     */
+    encode_text(text) {
+        const ptr0 = passStringToWasm0(text, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_encode_text(this.__wbg_ptr, ptr0, len0);
+        var v2 = getArrayU32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v2;
+    }
+    /**
+     * EOS (end of sequence) token ID from the GGUF model metadata, if present.
+     * Generation stops automatically when this token is produced.
+     * @returns {number | undefined}
+     */
+    get eos_token_id() {
+        const ret = wasm.flareengine_eos_token_id(this.__wbg_ptr);
+        return ret === 0x100000001 ? undefined : ret;
+    }
+    /**
+     * Streaming text-in / text-out generation with a per-token JS callback.
+     *
+     * Encodes `prompt` with the embedded GGUF vocabulary, generates up to
+     * `max_tokens` tokens, and calls `on_token(token_str)` with the decoded
+     * text for each token as it is produced.  Returns the number of tokens
+     * generated (excluding any EOS token).
+     *
+     * Returns 0 if no GGUF vocab is available.
+     *
+     * # Note on browser streaming
+     * `on_token` is called synchronously inside WASM, so the browser will
+     * not visually update between tokens.  For visible character-by-character
+     * output, use `begin_stream` + `next_token` with `requestAnimationFrame`.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * let out = '';
+     * const count = engine.generate_stream("What is Rust?", 128, (token) => {
+     *   out += token;
+     * });
+     * output.textContent = out;
+     * ```
+     * @param {string} prompt
+     * @param {number} max_tokens
+     * @param {Function} on_token
+     * @returns {number}
+     */
+    generate_stream(prompt, max_tokens, on_token) {
+        const ptr0 = passStringToWasm0(prompt, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_generate_stream(this.__wbg_ptr, ptr0, len0, max_tokens, on_token);
+        return ret >>> 0;
+    }
+    /**
+     * Streaming text-in / text-out with explicit sampling parameters.
+     *
+     * Like `generate_stream` but with the full set of sampling controls:
+     *
+     * - `temperature`: 0 = greedy, higher = more diverse
+     * - `top_p`: nucleus sampling (1.0 = disabled)
+     * - `top_k`: top-k sampling, applied when `top_p` is 1.0 and `min_p` is 0.0 (0 = disabled)
+     * - `repeat_penalty`: repetition penalty (1.0 = disabled, 1.1–1.3 = typical)
+     * - `min_p`: min-p threshold (0.0 = disabled)
+     *
+     * Encodes `prompt` with the embedded GGUF vocabulary, generates up to
+     * `max_tokens` tokens, and calls `on_token(token_str)` with the decoded
+     * text for each token.  Respects stop sequences registered via
+     * `add_stop_sequence`.  Returns the number of tokens generated.
+     *
+     * Returns 0 if no GGUF vocab is available.
+     *
+     * # JS example
+     * ```javascript
+     * engine.add_stop_sequence("<|im_end|>");
+     * engine.reset();
+     * let out = '';
+     * const count = engine.generate_stream_with_params(
+     *   prompt, 200, 0.8, 0.95, 40, 1.1, 0.0,
+     *   (token) => { out += token; }
+     * );
+     * ```
+     * @param {string} prompt
+     * @param {number} max_tokens
+     * @param {number} temperature
+     * @param {number} top_p
+     * @param {number} top_k
+     * @param {number} repeat_penalty
+     * @param {number} min_p
+     * @param {Function} on_token
+     * @returns {number}
+     */
+    generate_stream_with_params(prompt, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p, on_token) {
+        const ptr0 = passStringToWasm0(prompt, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_generate_stream_with_params(this.__wbg_ptr, ptr0, len0, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p, on_token);
+        return ret >>> 0;
+    }
+    /**
+     * Full text-in / text-out generation using the embedded GGUF vocabulary.
+     *
+     * Encodes `prompt` with the embedded vocab, runs greedy generation for up
+     * to `max_tokens` steps, then decodes the output back to text. Stops
+     * automatically at EOS.
+     *
+     * Returns an empty string if no GGUF vocab is available.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * const response = engine.generate_text("What is Rust?", 128);
+     * output.textContent = response;
+     * ```
+     * @param {string} prompt
+     * @param {number} max_tokens
+     * @returns {string}
+     */
+    generate_text(prompt, max_tokens) {
+        let deferred2_0;
+        let deferred2_1;
+        try {
+            const ptr0 = passStringToWasm0(prompt, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len0 = WASM_VECTOR_LEN;
+            const ret = wasm.flareengine_generate_text(this.__wbg_ptr, ptr0, len0, max_tokens);
+            deferred2_0 = ret[0];
+            deferred2_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred2_0, deferred2_1, 1);
+        }
+    }
+    /**
+     * Full text-in / text-out generation with explicit sampling parameters.
+     *
+     * Like `generate_text` but with the full set of sampling controls:
+     *
+     * - `temperature`: 0 = greedy, higher = more diverse
+     * - `top_p`: nucleus sampling (1.0 = disabled)
+     * - `top_k`: top-k sampling, applied when `top_p` is 1.0 and `min_p` is 0.0 (0 = disabled)
+     * - `repeat_penalty`: repetition penalty (1.0 = disabled)
+     * - `min_p`: min-p threshold (0.0 = disabled)
+     *
+     * Returns the decoded generated text. Returns an empty string if no GGUF vocab is available.
+     * Respects stop sequences registered via `add_stop_sequence`.
+     *
+     * # JS example
+     * ```javascript
+     * engine.reset();
+     * const response = engine.generate_text_with_params(
+     *   "What is Rust?", 128, 0.8, 0.95, 40, 1.1, 0.0
+     * );
+     * output.textContent = response;
+     * ```
+     * @param {string} prompt
+     * @param {number} max_tokens
+     * @param {number} temperature
+     * @param {number} top_p
+     * @param {number} top_k
+     * @param {number} repeat_penalty
+     * @param {number} min_p
+     * @returns {string}
+     */
+    generate_text_with_params(prompt, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p) {
+        let deferred2_0;
+        let deferred2_1;
+        try {
+            const ptr0 = passStringToWasm0(prompt, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len0 = WASM_VECTOR_LEN;
+            const ret = wasm.flareengine_generate_text_with_params(this.__wbg_ptr, ptr0, len0, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p);
+            deferred2_0 = ret[0];
+            deferred2_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred2_0, deferred2_1, 1);
+        }
+    }
+    /**
+     * Generate `max_tokens` tokens starting from `prompt_tokens` (greedy).
+     * Stops early at EOS. Returns a Uint32Array of generated token IDs.
+     * @param {Uint32Array} prompt_tokens
+     * @param {number} max_tokens
+     * @returns {Uint32Array}
+     */
+    generate_tokens(prompt_tokens, max_tokens) {
+        const ptr0 = passArray32ToWasm0(prompt_tokens, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_generate_tokens(this.__wbg_ptr, ptr0, len0, max_tokens);
+        var v2 = getArrayU32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v2;
+    }
+    /**
+     * Generate a batch of tokens with explicit sampling parameters.
+     *
+     * - `temperature`: 0 = greedy, higher = more diverse
+     * - `top_p`: nucleus sampling (1.0 = disabled)
+     * - `top_k`: top-k sampling, applied when `top_p` is 1.0 and `min_p` is 0.0 (0 = disabled)
+     * - `repeat_penalty`: repetition penalty applied to recently-seen tokens (1.0 = disabled)
+     * - `min_p`: min-p threshold (0.0 = disabled); applied after `top_p`, before `top_k`
+     *
+     * Stops early at EOS. Uses a fixed LCG RNG seed for reproducibility.
+     * @param {Uint32Array} prompt_tokens
+     * @param {number} max_tokens
+     * @param {number} temperature
+     * @param {number} top_p
+     * @param {number} top_k
+     * @param {number} repeat_penalty
+     * @param {number} min_p
+     * @returns {Uint32Array}
+     */
+    generate_with_params(prompt_tokens, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p) {
+        const ptr0 = passArray32ToWasm0(prompt_tokens, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_generate_with_params(this.__wbg_ptr, ptr0, len0, max_tokens, temperature, top_p, top_k, repeat_penalty, min_p);
+        var v2 = getArrayU32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v2;
+    }
+    /**
+     * Returns `true` if raw quantized weights are currently loaded.
+     * @returns {boolean}
+     */
+    get has_raw_weights() {
+        const ret = wasm.flareengine_has_raw_weights(this.__wbg_ptr);
+        return ret !== 0;
+    }
+    /**
+     * Get the hidden dimension.
+     * @returns {number}
+     */
+    get hidden_dim() {
+        const ret = wasm.flareengine_hidden_dim(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Try to initialise the WebGPU compute backend.
+     *
+     * Call this after `load()` to enable GPU-accelerated matrix operations
+     * (matvec, matmul, silu_mul). Falls back silently to CPU if WebGPU is
+     * unavailable or adapter request fails.
+     *
+     * Returns `true` if a GPU backend was successfully initialised.
+     *
+     * ```javascript
+     * const engine = FlareEngine.load(bytes);
+     * const gpuEnabled = await engine.init_gpu();
+     * console.log('GPU:', gpuEnabled);
+     * ```
+     * @returns {Promise<boolean>}
+     */
+    init_gpu() {
+        const ret = wasm.flareengine_init_gpu(this.__wbg_ptr);
+        return ret;
+    }
+    /**
+     * Initialise the WebGPU backend using previously serialised pipeline cache
+     * bytes (from `engine.pipeline_cache_data()`).
+     *
+     * On backends that support driver-managed pipeline caches (Vulkan native),
+     * this allows the driver to reuse compiled GPU machine code from a previous
+     * run, eliminating cold-start shader recompilation (typically 100ms–2s).
+     *
+     * On unsupported backends (WebGPU, Metal, DX12) this behaves identically to
+     * `init_gpu()` — the cache bytes are silently ignored.
+     *
+     * ```javascript
+     * const cached = localStorage.getItem('flare-pipeline-cache');
+     * const cacheBytes = cached ? new Uint8Array(JSON.parse(cached)) : new Uint8Array();
+     * await engine.init_gpu_with_cache(cacheBytes);
+     * // After inference, persist the cache:
+     * const data = engine.pipeline_cache_data();
+     * if (data.length > 0) {
+     *   localStorage.setItem('flare-pipeline-cache', JSON.stringify(Array.from(data)));
+     * }
+     * ```
+     * @param {Uint8Array} cache_data
+     * @returns {Promise<boolean>}
+     */
+    init_gpu_with_cache(cache_data) {
+        const ptr0 = passArray8ToWasm0(cache_data, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_init_gpu_with_cache(this.__wbg_ptr, ptr0, len0);
+        return ret;
+    }
+    /**
+     * Milliseconds spent in decode steps of the last generation call.
+     *
+     * For batch generation (`generate_tokens` etc.) this is always 0 — see
+     * `last_prefill_ms` for the total time.  For the streaming API this
+     * accumulates across all `next_token()` calls since the last
+     * `begin_stream()`.
+     * @returns {number}
+     */
+    get last_decode_ms() {
+        const ret = wasm.flareengine_last_decode_ms(this.__wbg_ptr);
+        return ret;
+    }
+    /**
+     * Raw pre-temperature logits from the most recent forward pass.
+     *
+     * Returns the full vocabulary logit vector as a `Float32Array`.  These
+     * are the raw values *before* temperature scaling, repetition penalty,
+     * or any sampling filter — equivalent to the model's raw next-token
+     * distribution.
+     *
+     * Useful for:
+     * - Scoring candidate continuations (classification, ranking)
+     * - Computing perplexity / cross-entropy
+     * - Inspecting the model's "confidence" about the next token
+     *
+     * Returns an empty array before any inference has been run, and is
+     * cleared by `reset()`.
+     *
+     * ```javascript
+     * engine.begin_stream(promptIds, 1); // one token prefill+decode
+     * engine.next_token();
+     * const logits = engine.last_logits; // Float32Array of vocab_size
+     * const topTokenId = logits.indexOf(Math.max(...logits));
+     * ```
+     * @returns {Float32Array}
+     */
+    get last_logits() {
+        const ret = wasm.flareengine_last_logits(this.__wbg_ptr);
+        var v1 = getArrayF32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v1;
+    }
+    /**
+     * Milliseconds spent in the last prefill (prompt processing) phase.
+     *
+     * For `generate_tokens` / `generate_text` / `generate_with_params` this
+     * covers the entire call (prefill + decode are not separated internally).
+     * For the streaming API (`begin_stream` + `next_token`) this covers only
+     * the `begin_stream()` call.
+     * @returns {number}
+     */
+    get last_prefill_ms() {
+        const ret = wasm.flareengine_last_prefill_ms(this.__wbg_ptr);
+        return ret;
+    }
+    /**
+     * Number of tokens generated by the last generation call (excludes prompt
+     * tokens and the EOS token itself).
+     * @returns {number}
+     */
+    get last_tokens_generated() {
+        const ret = wasm.flareengine_last_tokens_generated(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Load a GGUF model from a Uint8Array of bytes (e.g. from `fetch`).
+     * @param {Uint8Array} gguf_bytes
+     * @returns {FlareEngine}
+     */
+    static load(gguf_bytes) {
+        const ptr0 = passArray8ToWasm0(gguf_bytes, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_load(ptr0, len0);
+        if (ret[2]) {
+            throw takeFromExternrefTable0(ret[1]);
+        }
+        return FlareEngine.__wrap(ret[0]);
+    }
+    /**
+     * Load raw quantized weights from GGUF bytes so the GPU fused
+     * dequant+matvec kernels can be used during inference.
+     *
+     * Call this **after** `init_gpu()` so the backend is set before the raw
+     * weights are attached.  The method is a no-op (returns `false`) if a
+     * layer's weights are in an unsupported quantization format — the engine
+     * continues to work using the f32 path loaded at `FlareEngine.load()`.
+     *
+     * Returns `true` if all layers were loaded successfully, `false` if any
+     * layer fell back to the f32 path.
+     *
+     * ```javascript
+     * const engine = FlareEngine.load(bytes);
+     * await engine.init_gpu();
+     * const ok = engine.load_raw_weights(bytes);
+     * console.log('Raw weights loaded:', ok);
+     * ```
+     * @param {Uint8Array} gguf_bytes
+     * @returns {boolean}
+     */
+    load_raw_weights(gguf_bytes) {
+        const ptr0 = passArray8ToWasm0(gguf_bytes, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_load_raw_weights(this.__wbg_ptr, ptr0, len0);
+        return ret !== 0;
+    }
+    /**
+     * Maximum sequence length (context window size) of the loaded model.
+     *
+     * Use this to warn users when their prompt is approaching the limit.
+     * @returns {number}
+     */
+    get max_seq_len() {
+        const ret = wasm.flareengine_max_seq_len(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Merge a LoRA adapter (SafeTensors format) into the model weights.
+     *
+     * Pass the raw bytes of a `.safetensors` file containing LoRA A/B matrices.
+     * After merging, the adapter's effect is permanent for this engine instance;
+     * call `FlareEngine.load()` again to restore the base model.
+     *
+     * ```javascript
+     * const resp = await fetch('lora-adapter.safetensors');
+     * const bytes = new Uint8Array(await resp.arrayBuffer());
+     * engine.merge_lora(bytes);
+     * ```
+     * @param {Uint8Array} adapter_bytes
+     */
+    merge_lora(adapter_bytes) {
+        const ptr0 = passArray8ToWasm0(adapter_bytes, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_merge_lora(this.__wbg_ptr, ptr0, len0);
+        if (ret[1]) {
+            throw takeFromExternrefTable0(ret[0]);
+        }
+    }
+    /**
+     * Merge a LoRA adapter with a custom alpha scaling factor.
+     *
+     * Same as `merge_lora` but overrides the alpha value embedded in the
+     * adapter file.  The effective scaling is `alpha / rank`.
+     * @param {Uint8Array} adapter_bytes
+     * @param {number} alpha
+     */
+    merge_lora_with_alpha(adapter_bytes, alpha) {
+        const ptr0 = passArray8ToWasm0(adapter_bytes, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_merge_lora_with_alpha(this.__wbg_ptr, ptr0, len0, alpha);
+        if (ret[1]) {
+            throw takeFromExternrefTable0(ret[0]);
+        }
+    }
+    /**
+     * All GGUF model metadata as a JSON string.
+     *
+     * Returns a JSON object mapping each metadata key to its value.
+     * Large vocabulary arrays (`tokenizer.ggml.tokens`, `.merges`, `.scores`,
+     * `.added_tokens`) are omitted to keep the payload practical.
+     * Small arrays (≤ 64 entries) are included as JSON arrays.
+     *
+     * Returns `"{}"` if the model was not loaded from a GGUF file.
+     *
+     * ```javascript
+     * const meta = JSON.parse(engine.metadata_json);
+     * console.log(meta["llama.context_length"]); // e.g. 4096
+     * ```
+     * @returns {string}
+     */
+    get metadata_json() {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_metadata_json(this.__wbg_ptr);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Model display name from `general.name` in the GGUF metadata.
+     *
+     * Returns the human-readable name embedded by the model author (e.g.
+     * `"Llama 3.2 1B Instruct"`). Returns an empty string if the field is absent.
+     * @returns {string}
+     */
+    get model_name() {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_model_name(this.__wbg_ptr);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Generate and return the next token ID, or `undefined` when the stream
+     * is complete (EOS reached, `max_tokens` exhausted, or `stop_stream()`
+     * was called).
+     *
+     * Sampling parameters are those set by the most recent `begin_stream` or
+     * `begin_stream_with_params` call.  Call this inside
+     * `requestAnimationFrame` so the browser can update the DOM between
+     * tokens and the page remains responsive.
+     * @returns {number | undefined}
+     */
+    next_token() {
+        const ret = wasm.flareengine_next_token(this.__wbg_ptr);
+        return ret === 0x100000001 ? undefined : ret;
+    }
+    /**
+     * Get the number of attention heads.
+     * @returns {number}
+     */
+    get num_heads() {
+        const ret = wasm.flareengine_num_heads(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Get the number of layers.
+     * @returns {number}
+     */
+    get num_layers() {
+        const ret = wasm.flareengine_num_layers(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Apply final RMSNorm + output projection to a hidden state and
+     * return logits over the vocabulary.
+     *
+     * `hidden` must have length `hidden_dim`. The returned vector has
+     * length `vocab_size`.
+     * @param {Float32Array} hidden
+     * @returns {Float32Array}
+     */
+    output_projection(hidden) {
+        const ptr0 = passArrayF32ToWasm0(hidden, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareengine_output_projection(this.__wbg_ptr, ptr0, len0);
+        var v2 = getArrayF32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v2;
+    }
+    /**
+     * Return a JSON string summarising the performance metrics from the last
+     * generation call.
+     *
+     * ```javascript
+     * const perf = JSON.parse(engine.performance_summary());
+     * console.log(`TTFT: ${perf.prefill_ms.toFixed(1)} ms`);
+     * console.log(`Decode: ${perf.tokens_per_second.toFixed(1)} tok/s`);
+     * ```
+     * @returns {string}
+     */
+    performance_summary() {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_performance_summary(this.__wbg_ptr);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Serialise the driver-managed GPU pipeline cache to bytes.
+     *
+     * Returns an opaque blob that can be passed to `init_gpu_with_cache()` on
+     * the next startup to skip shader recompilation.  Store it in
+     * `localStorage` or `IndexedDB` between page loads.
+     *
+     * Returns an empty `Uint8Array` if no GPU is active, or if the current
+     * backend does not support pipeline caching (WebGPU, Metal, DX12).
+     * @returns {Uint8Array}
+     */
+    get pipeline_cache_data() {
+        const ret = wasm.flareengine_pipeline_cache_data(this.__wbg_ptr);
+        var v1 = getArrayU8FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 1, 1);
+        return v1;
+    }
+    /**
+     * Raw Jinja2 chat template string from the GGUF model metadata, if present.
+     *
+     * This is the `tokenizer.chat_template` field embedded by the model author.
+     * Use this with a JavaScript Jinja2 renderer (e.g. `nunjucks`) for accurate
+     * prompt formatting across all model families, rather than relying on the
+     * simplified built-in `apply_chat_template`.
+     *
+     * Returns `undefined` if the GGUF file did not include a chat template.
+     * @returns {string | undefined}
+     */
+    get raw_chat_template() {
+        const ret = wasm.flareengine_raw_chat_template(this.__wbg_ptr);
+        let v1;
+        if (ret[0] !== 0) {
+            v1 = getStringFromWasm0(ret[0], ret[1]).slice();
+            wasm.__wbindgen_free(ret[0], ret[1] * 1, 1);
+        }
+        return v1;
+    }
+    /**
+     * Current repetition-penalty window size (0 = disabled).
+     * @returns {number}
+     */
+    get repeat_last_n() {
+        const ret = wasm.flareengine_repeat_last_n(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Reset the KV cache (start a new conversation).
+     *
+     * Also clears stop sequences, the internal text accumulator, and
+     * restores the RNG seed to the default `0x12345678`.
+     */
+    reset() {
+        wasm.flareengine_reset(this.__wbg_ptr);
+    }
+    /**
+     * Set the repetition-penalty look-back window (number of recent tokens to
+     * penalise).  Use `0` to disable repetition penalty entirely.  Default: 64.
+     *
+     * Takes effect on the next `begin_stream*` call.
+     *
+     * # JS example
+     * ```javascript
+     * engine.set_repeat_last_n(128); // wider window for creative writing
+     * engine.set_repeat_last_n(0);   // disable repeat penalty
+     * ```
+     * @param {number} n
+     */
+    set_repeat_last_n(n) {
+        wasm.flareengine_set_repeat_last_n(this.__wbg_ptr, n);
+    }
+    /**
+     * Set the LCG RNG seed used for the next sampled generation call.
+     *
+     * Controls the random state passed to `begin_stream_with_params` and
+     * `generate_with_params`, enabling reproducible outputs.  The seed is
+     * applied on the next call and then *not* automatically reset, so the
+     * same seed will be reused on subsequent calls unless `set_rng_seed` or
+     * `reset()` is called again.
+     *
+     * `reset()` restores the seed to the default `0x12345678`.
+     *
+     * ```javascript
+     * engine.set_rng_seed(42);
+     * const out1 = engine.generate_text("Hello", 50);
+     * engine.set_rng_seed(42);
+     * const out2 = engine.generate_text("Hello", 50);
+     * // out1 === out2
+     * ```
+     * @param {number} seed
+     */
+    set_rng_seed(seed) {
+        wasm.flareengine_set_rng_seed(this.__wbg_ptr, seed);
+    }
+    /**
+     * Set how many top log-probability entries to capture after each forward
+     * pass.  Pass `0` (the default) to disable and save the computation.
+     *
+     * When enabled, `top_logprobs` is populated after every `next_token()`
+     * call and after every token in `generate_stream_with_params`.
+     *
+     * # JS example
+     * ```javascript
+     * engine.set_top_logprobs(5);
+     * engine.begin_stream(promptIds, 64);
+     * while (!engine.stream_done) {
+     *   engine.next_token();
+     *   const lp = engine.top_logprobs; // Float32Array [id0, lp0, id1, lp1, ...]
+     * }
+     * ```
+     * @param {number} n
+     */
+    set_top_logprobs(n) {
+        wasm.flareengine_set_top_logprobs(this.__wbg_ptr, n);
+    }
+    /**
+     * Signal the current stream to stop after the next `next_token()` call.
+     * The JS Stop button should call this, then wait for `next_token()` to
+     * return `undefined` before updating the UI.
+     */
+    stop_stream() {
+        wasm.flareengine_stop_stream(this.__wbg_ptr);
+    }
+    /**
+     * Whether the current stream has finished.
+     * @returns {boolean}
+     */
+    get stream_done() {
+        const ret = wasm.flareengine_stream_done(this.__wbg_ptr);
+        return ret !== 0;
+    }
+    /**
+     * Why the most-recent stream stopped.
+     *
+     * Returns one of:
+     * - `"eos"` — the model emitted the EOS token
+     * - `"length"` — `max_tokens` budget was exhausted
+     * - `"stop_sequence"` — a registered stop sequence was matched
+     * - `"user"` — `stop_stream()` was called
+     * - `""` (empty) — stream not yet started or still running
+     *
+     * # JS example
+     * ```javascript
+     * while (!engine.stream_done) engine.next_token();
+     * console.log("Stopped because:", engine.stream_stop_reason);
+     * ```
+     * @returns {string}
+     */
+    get stream_stop_reason() {
+        let deferred1_0;
+        let deferred1_1;
+        try {
+            const ret = wasm.flareengine_stream_stop_reason(this.__wbg_ptr);
+            deferred1_0 = ret[0];
+            deferred1_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+        }
+    }
+    /**
+     * Decode throughput in tokens per second for the last generation call.
+     *
+     * For the streaming API this is calculated from `last_decode_ms`.
+     * For batch generation this is calculated from `last_prefill_ms`
+     * (the total call duration).
+     *
+     * Returns 0.0 if no generation has been run or if timing data is
+     * unavailable.
+     * @returns {number}
+     */
+    get tokens_per_second() {
+        const ret = wasm.flareengine_tokens_per_second(this.__wbg_ptr);
+        return ret;
+    }
+    /**
+     * How many tokens of context space remain before the window is full.
+     *
+     * Equivalent to `max_seq_len - tokens_used`. Returns 0 when the context is
+     * already full or `max_seq_len` is 0.
+     *
+     * # JS example
+     * ```javascript
+     * if (engine.tokens_remaining < 64) {
+     *   console.warn("Context window almost full — consider resetting.");
+     * }
+     * ```
+     * @returns {number}
+     */
+    get tokens_remaining() {
+        const ret = wasm.flareengine_tokens_remaining(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Number of tokens currently consumed in the KV-cache session (prompt + generated).
+     *
+     * Updated after every generation call; reset to 0 by `engine.reset()`.
+     * Use with `max_seq_len` to build a context-usage progress bar.
+     * @returns {number}
+     */
+    get tokens_used() {
+        const ret = wasm.flareengine_tokens_used(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Interleaved top-N log-probabilities from the last forward pass.
+     *
+     * Layout: `[token_id_0 as f32, log_prob_0, token_id_1 as f32, log_prob_1, ...]`
+     * sorted by descending log-probability.  Length is `top_logprobs_n * 2`.
+     *
+     * Returns an empty array if `set_top_logprobs(0)` (default) or before
+     * any inference has been run.
+     * @returns {Float32Array}
+     */
+    get top_logprobs() {
+        const ret = wasm.flareengine_top_logprobs(this.__wbg_ptr);
+        var v1 = getArrayF32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v1;
+    }
+    /**
+     * Truncate `text` so that it fits within `budget` tokens when encoded.
+     *
+     * Encodes `text` with the embedded GGUF vocabulary, keeps the **last**
+     * `budget` tokens (tail of the text is preferred, so recent context is
+     * preserved), and decodes them back to a string.  Returns `text` unchanged
+     * if it already fits or if no vocab is available.
+     *
+     * A typical call reserves space for the system prompt + generated output:
+     *
+     * ```javascript
+     * // Keep only the tail of the conversation that fits in the context
+     * const budget = engine.max_seq_len - 256; // leave 256 tokens for output
+     * const trimmed = engine.truncate_to_context(conversationText, budget);
+     * ```
+     * @param {string} text
+     * @param {number} budget
+     * @returns {string}
+     */
+    truncate_to_context(text, budget) {
+        let deferred2_0;
+        let deferred2_1;
+        try {
+            const ptr0 = passStringToWasm0(text, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len0 = WASM_VECTOR_LEN;
+            const ret = wasm.flareengine_truncate_to_context(this.__wbg_ptr, ptr0, len0, budget);
+            deferred2_0 = ret[0];
+            deferred2_1 = ret[1];
+            return getStringFromWasm0(ret[0], ret[1]);
+        } finally {
+            wasm.__wbindgen_free(deferred2_0, deferred2_1, 1);
+        }
+    }
+    /**
+     * Get the vocabulary size of the loaded model.
+     * @returns {number}
+     */
+    get vocab_size() {
+        const ret = wasm.flareengine_vocab_size(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Run a single dummy forward pass to pre-compile WebGPU shader pipelines.
+     *
+     * WebGPU (and wgpu on native) compiles shader pipelines lazily on the
+     * first dispatch.  This causes a noticeable latency spike — often 100ms
+     * to several seconds — when the user makes their first inference request.
+     *
+     * Call `warmup()` once after `init_gpu()` completes to trigger all shader
+     * compilations in the background so the first real request feels fast.
+     * The KV cache is reset after the warmup so the engine is in a clean state.
+     *
+     * Returns `true` if the warmup forward pass ran without error, `false` if
+     * the model has not been loaded.
+     *
+     * # JS example
+     * ```javascript
+     * const engine = FlareEngine.load(bytes);
+     * await engine.init_gpu();
+     * engine.warmup(); // trigger shader compilation
+     * // First real inference is now fast
+     * engine.begin_stream(promptIds, 128);
+     * ```
+     * @returns {boolean}
+     */
+    warmup() {
+        const ret = wasm.flareengine_warmup(this.__wbg_ptr);
+        return ret !== 0;
+    }
+}
+if (Symbol.dispose) FlareEngine.prototype[Symbol.dispose] = FlareEngine.prototype.free;
+
+/**
+ * Progressive loader that fetches a GGUF model from a URL with streaming
+ * download progress.
+ *
+ * This enables the browser demo to show download progress as the model
+ * arrives over the network, then layer-loading progress as the model is
+ * parsed. For a 500MB Q4 model the download phase dominates; displaying
+ * progress prevents the page from appearing frozen.
+ *
+ * # JS example
+ *
+ * ```javascript
+ * const loader = new FlareProgressiveLoader('https://example.com/model.gguf');
+ * const engine = await loader.load((loaded, total) => {
+ *   const pct = total > 0 ? Math.round(loaded / total * 100) : 0;
+ *   progressBar.value = pct / 100;
+ *   statusText.textContent = `Downloading… ${pct}%`;
+ * });
+ * ```
+ */
+export class FlareProgressiveLoader {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        FlareProgressiveLoaderFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_flareprogressiveloader_free(ptr, 0);
+    }
+    /**
+     * Fetch the model from the URL, calling `on_progress(loaded_bytes, total_bytes)`
+     * as each chunk arrives, then parse and return a `FlareEngine`.
+     *
+     * `total_bytes` is 0 when the server does not send a `Content-Length` header
+     * (e.g. when the response is gzip-compressed or chunked).
+     * @param {Function} on_progress
+     * @returns {Promise<FlareEngine>}
+     */
+    load(on_progress) {
+        const ret = wasm.flareprogressiveloader_load(this.__wbg_ptr, on_progress);
+        return ret;
+    }
+    /**
+     * Create a loader for the given model URL.
+     * @param {string} url
+     */
+    constructor(url) {
+        const ptr0 = passStringToWasm0(url, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flareprogressiveloader_new(ptr0, len0);
+        this.__wbg_ptr = ret >>> 0;
+        FlareProgressiveLoaderFinalization.register(this, this.__wbg_ptr, this);
+        return this;
+    }
+}
+if (Symbol.dispose) FlareProgressiveLoader.prototype[Symbol.dispose] = FlareProgressiveLoader.prototype.free;
+
+/**
+ * BPE tokenizer exported to JS for encoding prompts and decoding generated tokens.
+ *
+ * Load from a HuggingFace `tokenizer.json` string, then use `encode` / `decode`
+ * in coordination with `FlareEngine` to run full text-in / text-out inference.
+ *
+ * # JS example
+ *
+ * ```javascript
+ * const resp = await fetch('tokenizer.json');
+ * const json = await resp.text();
+ * const tok = FlareTokenizer.from_json(json);
+ *
+ * const ids = tok.encode("Hello, world!");
+ * const engine = FlareEngine.load(modelBytes);
+ * const out = engine.generate_tokens(ids, 64);
+ * console.log(tok.decode(out));
+ * ```
+ */
+export class FlareTokenizer {
+    static __wrap(ptr) {
+        ptr = ptr >>> 0;
+        const obj = Object.create(FlareTokenizer.prototype);
+        obj.__wbg_ptr = ptr;
+        FlareTokenizerFinalization.register(obj, obj.__wbg_ptr, obj);
+        return obj;
+    }
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        FlareTokenizerFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_flaretokenizer_free(ptr, 0);
+    }
+    /**
+     * BOS (beginning of sequence) token ID, if defined.
+     * @returns {number | undefined}
+     */
+    get bos_token_id() {
+        const ret = wasm.flaretokenizer_bos_token_id(this.__wbg_ptr);
+        return ret === 0x100000001 ? undefined : ret;
+    }
+    /**
+     * Decode a sequence of token IDs to text.
+     * @param {Uint32Array} tokens
+     * @returns {string}
+     */
+    decode(tokens) {
+        let deferred3_0;
+        let deferred3_1;
+        try {
+            const ptr0 = passArray32ToWasm0(tokens, wasm.__wbindgen_malloc);
+            const len0 = WASM_VECTOR_LEN;
+            const ret = wasm.flaretokenizer_decode(this.__wbg_ptr, ptr0, len0);
+            var ptr2 = ret[0];
+            var len2 = ret[1];
+            if (ret[3]) {
+                ptr2 = 0; len2 = 0;
+                throw takeFromExternrefTable0(ret[2]);
+            }
+            deferred3_0 = ptr2;
+            deferred3_1 = len2;
+            return getStringFromWasm0(ptr2, len2);
+        } finally {
+            wasm.__wbindgen_free(deferred3_0, deferred3_1, 1);
+        }
+    }
+    /**
+     * Decode a single token ID to text (useful for streaming output).
+     * @param {number} token_id
+     * @returns {string}
+     */
+    decode_one(token_id) {
+        let deferred2_0;
+        let deferred2_1;
+        try {
+            const ret = wasm.flaretokenizer_decode_one(this.__wbg_ptr, token_id);
+            var ptr1 = ret[0];
+            var len1 = ret[1];
+            if (ret[3]) {
+                ptr1 = 0; len1 = 0;
+                throw takeFromExternrefTable0(ret[2]);
+            }
+            deferred2_0 = ptr1;
+            deferred2_1 = len1;
+            return getStringFromWasm0(ptr1, len1);
+        } finally {
+            wasm.__wbindgen_free(deferred2_0, deferred2_1, 1);
+        }
+    }
+    /**
+     * Encode text to a sequence of token IDs.
+     * @param {string} text
+     * @returns {Uint32Array}
+     */
+    encode(text) {
+        const ptr0 = passStringToWasm0(text, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flaretokenizer_encode(this.__wbg_ptr, ptr0, len0);
+        if (ret[3]) {
+            throw takeFromExternrefTable0(ret[2]);
+        }
+        var v2 = getArrayU32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v2;
+    }
+    /**
+     * EOS (end of sequence) token ID, if defined.
+     * @returns {number | undefined}
+     */
+    get eos_token_id() {
+        const ret = wasm.flaretokenizer_eos_token_id(this.__wbg_ptr);
+        return ret === 0x100000001 ? undefined : ret;
+    }
+    /**
+     * Load a tokenizer from the text of a HuggingFace `tokenizer.json` file.
+     * @param {string} json
+     * @returns {FlareTokenizer}
+     */
+    static from_json(json) {
+        const ptr0 = passStringToWasm0(json, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.flaretokenizer_from_json(ptr0, len0);
+        if (ret[2]) {
+            throw takeFromExternrefTable0(ret[1]);
+        }
+        return FlareTokenizer.__wrap(ret[0]);
+    }
+    /**
+     * Vocabulary size.
+     * @returns {number}
+     */
+    get vocab_size() {
+        const ret = wasm.flaretokenizer_vocab_size(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+}
+if (Symbol.dispose) FlareTokenizer.prototype[Symbol.dispose] = FlareTokenizer.prototype.free;
+
+/**
+ * Save model bytes to OPFS.
+ *
+ * Creates the `flare-models` directory if it does not exist.  Overwrites any
+ * existing file with the same name.
+ * @param {string} model_name
+ * @param {Uint8Array} data
+ * @returns {Promise<void>}
+ */
+export function cache_model(model_name, data) {
+    const ptr0 = passStringToWasm0(model_name, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ptr1 = passArray8ToWasm0(data, wasm.__wbindgen_malloc);
+    const len1 = WASM_VECTOR_LEN;
+    const ret = wasm.cache_model(ptr0, len0, ptr1, len1);
+    return ret;
+}
+
+/**
+ * Delete a cached model from OPFS.
+ * @param {string} model_name
+ * @returns {Promise<void>}
+ */
+export function delete_cached_model(model_name) {
+    const ptr0 = passStringToWasm0(model_name, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.delete_cached_model(ptr0, len0);
+    return ret;
+}
+
+/**
+ * Get basic device info as a JSON string.
+ * @returns {string}
+ */
+export function device_info() {
+    let deferred1_0;
+    let deferred1_1;
+    try {
+        const ret = wasm.device_info();
+        deferred1_0 = ret[0];
+        deferred1_1 = ret[1];
+        return getStringFromWasm0(ret[0], ret[1]);
+    } finally {
+        wasm.__wbindgen_free(deferred1_0, deferred1_1, 1);
+    }
+}
+
+/**
+ * Check if a model is cached in OPFS by name.
+ *
+ * Returns `false` if OPFS is unavailable or the model is not found.
+ * @param {string} model_name
+ * @returns {Promise<boolean>}
+ */
+export function is_model_cached(model_name) {
+    const ptr0 = passStringToWasm0(model_name, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.is_model_cached(ptr0, len0);
+    return ret;
+}
+
+/**
+ * List all cached models with their sizes (in bytes).
+ *
+ * Returns a JSON-serialised array of objects: `[{name: string, size: number}, ...]`.
+ * Returns `"[]"` if OPFS is unavailable or the models directory does not exist.
+ * @returns {Promise<any>}
+ */
+export function list_cached_models() {
+    const ret = wasm.list_cached_models();
+    return ret;
+}
+
+/**
+ * Load model bytes from OPFS.
+ *
+ * Returns `null` (JS) / `None` (Rust) if the model is not cached or OPFS is
+ * unavailable.
+ * @param {string} model_name
+ * @returns {Promise<any>}
+ */
+export function load_cached_model(model_name) {
+    const ptr0 = passStringToWasm0(model_name, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.load_cached_model(ptr0, len0);
+    return ret;
+}
+
+/**
+ * Set up better panic messages in the browser console.
+ */
+export function start() {
+    wasm.start();
+}
+
+/**
+ * Get storage usage and quota estimate.
+ *
+ * Returns a JSON string: `{usage: number, quota: number}`.
+ * Returns `"{}"` if the Storage API is unavailable.
+ * @returns {Promise<any>}
+ */
+export function storage_estimate() {
+    const ret = wasm.storage_estimate();
+    return ret;
+}
+
+/**
+ * Check if this WASM build was compiled with relaxed SIMD support.
+ *
+ * Relaxed SIMD provides hardware-specific faster operations like fused
+ * multiply-add (`f32x4_relaxed_madd`) that map directly to ARM NEON and
+ * x86 SSE/AVX FMA instructions. When enabled, matvec operations use FMA
+ * for ~15-30% speedup.
+ *
+ * This is a compile-time feature: the WASM binary either includes relaxed
+ * SIMD instructions or it does not. The browser validates them at module
+ * load time, so if this module loaded successfully and returns `true`,
+ * relaxed SIMD is active.
+ * @returns {boolean}
+ */
+export function supports_relaxed_simd() {
+    const ret = wasm.supports_relaxed_simd();
+    return ret !== 0;
+}
+
+/**
+ * Check if the browser exposes the Web Speech API for speech recognition.
+ *
+ * This probes `window.SpeechRecognition` and the WebKit-prefixed
+ * `window.webkitSpeechRecognition`. Returning `true` means the demo voice
+ * mode can capture microphone input and produce transcripts through the
+ * platform speech engine. This is a foundation for the voice pipeline
+ * (issue #395); a fully offline path will eventually run Whisper in WASM.
+ * @returns {boolean}
+ */
+export function supports_speech_recognition() {
+    const ret = wasm.supports_speech_recognition();
+    return ret !== 0;
+}
+
+/**
+ * Check if the browser exposes the Web Speech API for speech synthesis.
+ *
+ * Returns `true` when `window.speechSynthesis` is available, enabling the
+ * demo voice mode to speak model responses. A fully offline path will
+ * eventually run a neural TTS model in WASM.
+ * @returns {boolean}
+ */
+export function supports_speech_synthesis() {
+    const ret = wasm.supports_speech_synthesis();
+    return ret !== 0;
+}
+
+/**
+ * Check if WebNN is available in the current browser.
+ *
+ * WebNN (`navigator.ml`) exposes neural-network acceleration through
+ * platform NPUs/DSPs. This is a foundation check so JS code can decide
+ * whether to build a WebNN graph from exported weights.
+ * @returns {boolean}
+ */
+export function supports_webnn() {
+    const ret = wasm.supports_webnn();
+    return ret !== 0;
+}
+
+/**
+ * Check if WebTransport is available in the current browser.
+ *
+ * WebTransport (`window.WebTransport`) is a modern transport API built on
+ * HTTP/3 QUIC streams. It allows opening multiple parallel bidirectional
+ * streams to the same origin with lower head-of-line blocking than fetch().
+ * Useful for progressive model loading where different byte ranges of the
+ * GGUF file can be downloaded concurrently.
+ *
+ * Note: actually using WebTransport for parallel range downloads requires
+ * server-side support (HTTP/3 endpoint that accepts byte-range requests
+ * on streams). This check only reports browser capability — the JS loader
+ * will fall back to `fetch()` when the server does not cooperate.
+ * @returns {boolean}
+ */
+export function supports_webtransport() {
+    const ret = wasm.supports_webtransport();
+    return ret !== 0;
+}
+
+/**
+ * Check if WebGPU is available in the current browser.
+ * @returns {boolean}
+ */
+export function webgpu_available() {
+    const ret = wasm.webgpu_available();
+    return ret !== 0;
+}
+
+function __wbg_get_imports() {
+    const import0 = {
+        __proto__: null,
+        __wbg_Error_2e59b1b37a9a34c3: function(arg0, arg1) {
+            const ret = Error(getStringFromWasm0(arg0, arg1));
+            return ret;
+        },
+        __wbg_Window_412fe051c1aa1519: function(arg0) {
+            const ret = arg0.Window;
+            return ret;
+        },
+        __wbg_WorkerGlobalScope_349300f9b277afe1: function(arg0) {
+            const ret = arg0.WorkerGlobalScope;
+            return ret;
+        },
+        __wbg___wbindgen_boolean_get_a86c216575a75c30: function(arg0) {
+            const v = arg0;
+            const ret = typeof(v) === 'boolean' ? v : undefined;
+            return isLikeNone(ret) ? 0xFFFFFF : ret ? 1 : 0;
+        },
+        __wbg___wbindgen_debug_string_dd5d2d07ce9e6c57: function(arg0, arg1) {
+            const ret = debugString(arg1);
+            const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        },
+        __wbg___wbindgen_is_function_49868bde5eb1e745: function(arg0) {
+            const ret = typeof(arg0) === 'function';
+            return ret;
+        },
+        __wbg___wbindgen_is_null_344c8750a8525473: function(arg0) {
+            const ret = arg0 === null;
+            return ret;
+        },
+        __wbg___wbindgen_is_undefined_c0cca72b82b86f4d: function(arg0) {
+            const ret = arg0 === undefined;
+            return ret;
+        },
+        __wbg___wbindgen_string_get_914df97fcfa788f2: function(arg0, arg1) {
+            const obj = arg1;
+            const ret = typeof(obj) === 'string' ? obj : undefined;
+            var ptr1 = isLikeNone(ret) ? 0 : passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            var len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        },
+        __wbg___wbindgen_throw_81fc77679af83bc6: function(arg0, arg1) {
+            throw new Error(getStringFromWasm0(arg0, arg1));
+        },
+        __wbg__wbg_cb_unref_3c3b4f651835fbcb: function(arg0) {
+            arg0._wbg_cb_unref();
+        },
+        __wbg_arrayBuffer_7bba74066875530e: function(arg0) {
+            const ret = arg0.arrayBuffer();
+            return ret;
+        },
+        __wbg_beginComputePass_097033d61ef8af0f: function(arg0, arg1) {
+            const ret = arg0.beginComputePass(arg1);
+            return ret;
+        },
+        __wbg_body_9a25d64338506fbe: function(arg0) {
+            const ret = arg0.body;
+            return isLikeNone(ret) ? 0 : addToExternrefTable0(ret);
+        },
+        __wbg_buffer_a77cc90da4bdb503: function(arg0) {
+            const ret = arg0.buffer;
+            return ret;
+        },
+        __wbg_call_368fa9c372d473ba: function() { return handleError(function (arg0, arg1, arg2, arg3) {
+            const ret = arg0.call(arg1, arg2, arg3);
+            return ret;
+        }, arguments); },
+        __wbg_call_d578befcc3145dee: function() { return handleError(function (arg0, arg1, arg2) {
+            const ret = arg0.call(arg1, arg2);
+            return ret;
+        }, arguments); },
+        __wbg_close_37e34297940956fd: function(arg0) {
+            const ret = arg0.close();
+            return ret;
+        },
+        __wbg_copyBufferToBuffer_99ba10ae51f20b8a: function() { return handleError(function (arg0, arg1, arg2, arg3, arg4, arg5) {
+            arg0.copyBufferToBuffer(arg1, arg2, arg3, arg4, arg5);
+        }, arguments); },
+        __wbg_createBindGroupLayout_1d37ac0dabfbed28: function() { return handleError(function (arg0, arg1) {
+            const ret = arg0.createBindGroupLayout(arg1);
+            return ret;
+        }, arguments); },
+        __wbg_createBindGroup_3bccbd7517f0708e: function(arg0, arg1) {
+            const ret = arg0.createBindGroup(arg1);
+            return ret;
+        },
+        __wbg_createBuffer_24b346170c9f54c8: function() { return handleError(function (arg0, arg1) {
+            const ret = arg0.createBuffer(arg1);
+            return ret;
+        }, arguments); },
+        __wbg_createCommandEncoder_48a406baaa084912: function(arg0, arg1) {
+            const ret = arg0.createCommandEncoder(arg1);
+            return ret;
+        },
+        __wbg_createComputePipeline_4efb4ca205a4b557: function(arg0, arg1) {
+            const ret = arg0.createComputePipeline(arg1);
+            return ret;
+        },
+        __wbg_createPipelineLayout_f668b6fbdf877ab3: function(arg0, arg1) {
+            const ret = arg0.createPipelineLayout(arg1);
+            return ret;
+        },
+        __wbg_createShaderModule_1b0812f3a4503221: function(arg0, arg1) {
+            const ret = arg0.createShaderModule(arg1);
+            return ret;
+        },
+        __wbg_createWritable_d5314165379c13be: function(arg0) {
+            const ret = arg0.createWritable();
+            return ret;
+        },
+        __wbg_dispatchWorkgroups_1b750cb68e2eb693: function(arg0, arg1, arg2, arg3) {
+            arg0.dispatchWorkgroups(arg1 >>> 0, arg2 >>> 0, arg3 >>> 0);
+        },
+        __wbg_end_fd65a01a19361ec7: function(arg0) {
+            arg0.end();
+        },
+        __wbg_entries_69b67e91e74ba327: function(arg0) {
+            const ret = arg0.entries();
+            return ret;
+        },
+        __wbg_error_a6fa202b58aa1cd3: function(arg0, arg1) {
+            let deferred0_0;
+            let deferred0_1;
+            try {
+                deferred0_0 = arg0;
+                deferred0_1 = arg1;
+                console.error(getStringFromWasm0(arg0, arg1));
+            } finally {
+                wasm.__wbindgen_free(deferred0_0, deferred0_1, 1);
+            }
+        },
+        __wbg_estimate_790345f187f17044: function() { return handleError(function (arg0) {
+            const ret = arg0.estimate();
+            return ret;
+        }, arguments); },
+        __wbg_eval_db8671e4e6469929: function() { return handleError(function (arg0, arg1) {
+            const ret = eval(getStringFromWasm0(arg0, arg1));
+            return ret;
+        }, arguments); },
+        __wbg_features_205df3dd891b74bf: function(arg0) {
+            const ret = arg0.features;
+            return ret;
+        },
+        __wbg_features_efadd23951712b29: function(arg0) {
+            const ret = arg0.features;
+            return ret;
+        },
+        __wbg_fetch_ca19a9480623b9a8: function(arg0, arg1, arg2) {
+            const ret = arg0.fetch(getStringFromWasm0(arg1, arg2));
+            return ret;
+        },
+        __wbg_finish_2440fb64e53f7d5a: function(arg0, arg1) {
+            const ret = arg0.finish(arg1);
+            return ret;
+        },
+        __wbg_finish_4b40810f0b577bc2: function(arg0) {
+            const ret = arg0.finish();
+            return ret;
+        },
+        __wbg_flareengine_new: function(arg0) {
+            const ret = FlareEngine.__wrap(arg0);
+            return ret;
+        },
+        __wbg_from_741da0f916ab74aa: function(arg0) {
+            const ret = Array.from(arg0);
+            return ret;
+        },
+        __wbg_getDirectoryHandle_a38f7b2c1aa52af4: function(arg0, arg1, arg2, arg3) {
+            const ret = arg0.getDirectoryHandle(getStringFromWasm0(arg1, arg2), arg3);
+            return ret;
+        },
+        __wbg_getDirectory_3af764c18446017f: function(arg0) {
+            const ret = arg0.getDirectory();
+            return ret;
+        },
+        __wbg_getFileHandle_029e7a3c6dee72cb: function(arg0, arg1, arg2) {
+            const ret = arg0.getFileHandle(getStringFromWasm0(arg1, arg2));
+            return ret;
+        },
+        __wbg_getFileHandle_326ca47811ae37a1: function(arg0, arg1, arg2, arg3) {
+            const ret = arg0.getFileHandle(getStringFromWasm0(arg1, arg2), arg3);
+            return ret;
+        },
+        __wbg_getFile_0e25dfe508c6bd0a: function(arg0) {
+            const ret = arg0.getFile();
+            return ret;
+        },
+        __wbg_getMappedRange_55878eb97535ca19: function() { return handleError(function (arg0, arg1, arg2) {
+            const ret = arg0.getMappedRange(arg1, arg2);
+            return ret;
+        }, arguments); },
+        __wbg_getReader_3bcb712b2f3b80aa: function(arg0) {
+            const ret = arg0.getReader();
+            return ret;
+        },
+        __wbg_get_4848e350b40afc16: function(arg0, arg1) {
+            const ret = arg0[arg1 >>> 0];
+            return ret;
+        },
+        __wbg_get_5caaa5a9aae7e0b1: function() { return handleError(function (arg0, arg1, arg2, arg3) {
+            const ret = arg1.get(getStringFromWasm0(arg2, arg3));
+            var ptr1 = isLikeNone(ret) ? 0 : passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            var len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        }, arguments); },
+        __wbg_get_f96702c6245e4ef9: function() { return handleError(function (arg0, arg1) {
+            const ret = Reflect.get(arg0, arg1);
+            return ret;
+        }, arguments); },
+        __wbg_get_quota_945897ba0f160371: function(arg0, arg1) {
+            const ret = arg1.quota;
+            getDataViewMemory0().setFloat64(arg0 + 8 * 1, isLikeNone(ret) ? 0 : ret, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, !isLikeNone(ret), true);
+        },
+        __wbg_get_usage_2ea0330cfaeab2c1: function(arg0, arg1) {
+            const ret = arg1.usage;
+            getDataViewMemory0().setFloat64(arg0 + 8 * 1, isLikeNone(ret) ? 0 : ret, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, !isLikeNone(ret), true);
+        },
+        __wbg_gpu_bafbc1407fe850fb: function(arg0) {
+            const ret = arg0.gpu;
+            return ret;
+        },
+        __wbg_has_dc80aa6186153231: function(arg0, arg1, arg2) {
+            const ret = arg0.has(getStringFromWasm0(arg1, arg2));
+            return ret;
+        },
+        __wbg_headers_e08dcb5aa09b9a63: function(arg0) {
+            const ret = arg0.headers;
+            return ret;
+        },
+        __wbg_instanceof_GpuAdapter_aff4b0f95a6c1c3e: function(arg0) {
+            let result;
+            try {
+                result = arg0 instanceof GPUAdapter;
+            } catch (_) {
+                result = false;
+            }
+            const ret = result;
+            return ret;
+        },
+        __wbg_instanceof_ReadableStreamDefaultReader_10d3f15a012b70d7: function(arg0) {
+            let result;
+            try {
+                result = arg0 instanceof ReadableStreamDefaultReader;
+            } catch (_) {
+                result = false;
+            }
+            const ret = result;
+            return ret;
+        },
+        __wbg_instanceof_Response_06795eab66cc4036: function(arg0) {
+            let result;
+            try {
+                result = arg0 instanceof Response;
+            } catch (_) {
+                result = false;
+            }
+            const ret = result;
+            return ret;
+        },
+        __wbg_instanceof_Window_c0fee4c064502536: function(arg0) {
+            let result;
+            try {
+                result = arg0 instanceof Window;
+            } catch (_) {
+                result = false;
+            }
+            const ret = result;
+            return ret;
+        },
+        __wbg_label_4b6427d9045e3926: function(arg0, arg1) {
+            const ret = arg1.label;
+            const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        },
+        __wbg_length_0c32cb8543c8e4c8: function(arg0) {
+            const ret = arg0.length;
+            return ret;
+        },
+        __wbg_mapAsync_f7fe2e4825742580: function(arg0, arg1, arg2, arg3) {
+            const ret = arg0.mapAsync(arg1 >>> 0, arg2, arg3);
+            return ret;
+        },
+        __wbg_navigator_9b09ea705d03d227: function(arg0) {
+            const ret = arg0.navigator;
+            return ret;
+        },
+        __wbg_navigator_af52153252bdf29d: function(arg0) {
+            const ret = arg0.navigator;
+            return ret;
+        },
+        __wbg_new_227d7c05414eb861: function() {
+            const ret = new Error();
+            return ret;
+        },
+        __wbg_new_4f9fafbb3909af72: function() {
+            const ret = new Object();
+            return ret;
+        },
+        __wbg_new_a560378ea1240b14: function(arg0) {
+            const ret = new Uint8Array(arg0);
+            return ret;
+        },
+        __wbg_new_f3c9df4f38f3f798: function() {
+            const ret = new Array();
+            return ret;
+        },
+        __wbg_new_from_slice_2580ff33d0d10520: function(arg0, arg1) {
+            const ret = new Uint8Array(getArrayU8FromWasm0(arg0, arg1));
+            return ret;
+        },
+        __wbg_new_typed_14d7cc391ce53d2c: function(arg0, arg1) {
+            try {
+                var state0 = {a: arg0, b: arg1};
+                var cb0 = (arg0, arg1) => {
+                    const a = state0.a;
+                    state0.a = 0;
+                    try {
+                        return wasm_bindgen__convert__closures_____invoke__hcdfd434894ba1863(a, state0.b, arg0, arg1);
+                    } finally {
+                        state0.a = a;
+                    }
+                };
+                const ret = new Promise(cb0);
+                return ret;
+            } finally {
+                state0.a = 0;
+            }
+        },
+        __wbg_new_with_byte_offset_and_length_6bfc75833d6170c8: function(arg0, arg1, arg2) {
+            const ret = new Uint8Array(arg0, arg1 >>> 0, arg2 >>> 0);
+            return ret;
+        },
+        __wbg_next_072e78dcf497124d: function() { return handleError(function (arg0) {
+            const ret = arg0.next();
+            return ret;
+        }, arguments); },
+        __wbg_now_2c44418ca0623664: function(arg0) {
+            const ret = arg0.now();
+            return ret;
+        },
+        __wbg_ok_36f7b13b74596c24: function(arg0) {
+            const ret = arg0.ok;
+            return ret;
+        },
+        __wbg_performance_5ed3f6a3bbe36d0d: function(arg0) {
+            const ret = arg0.performance;
+            return isLikeNone(ret) ? 0 : addToExternrefTable0(ret);
+        },
+        __wbg_prototypesetcall_3e05eb9545565046: function(arg0, arg1, arg2) {
+            Uint8Array.prototype.set.call(getArrayU8FromWasm0(arg0, arg1), arg2);
+        },
+        __wbg_push_6bdbc990be5ac37b: function(arg0, arg1) {
+            const ret = arg0.push(arg1);
+            return ret;
+        },
+        __wbg_queueMicrotask_abaf92f0bd4e80a4: function(arg0) {
+            const ret = arg0.queueMicrotask;
+            return ret;
+        },
+        __wbg_queueMicrotask_df5a6dac26d818f3: function(arg0) {
+            queueMicrotask(arg0);
+        },
+        __wbg_queue_3e40156d83b9183e: function(arg0) {
+            const ret = arg0.queue;
+            return ret;
+        },
+        __wbg_read_316bf844c93a6ccc: function(arg0) {
+            const ret = arg0.read();
+            return ret;
+        },
+        __wbg_removeEntry_f038ab74448d1824: function(arg0, arg1, arg2) {
+            const ret = arg0.removeEntry(getStringFromWasm0(arg1, arg2));
+            return ret;
+        },
+        __wbg_requestAdapter_245da40985c2fdc5: function(arg0, arg1) {
+            const ret = arg0.requestAdapter(arg1);
+            return ret;
+        },
+        __wbg_requestDevice_28434913a23418c4: function(arg0, arg1) {
+            const ret = arg0.requestDevice(arg1);
+            return ret;
+        },
+        __wbg_resolve_0a79de24e9d2267b: function(arg0) {
+            const ret = Promise.resolve(arg0);
+            return ret;
+        },
+        __wbg_setBindGroup_98f0303f15c3cfb4: function() { return handleError(function (arg0, arg1, arg2, arg3, arg4, arg5, arg6) {
+            arg0.setBindGroup(arg1 >>> 0, arg2, getArrayU32FromWasm0(arg3, arg4), arg5, arg6 >>> 0);
+        }, arguments); },
+        __wbg_setBindGroup_bc67abae8c962082: function(arg0, arg1, arg2) {
+            arg0.setBindGroup(arg1 >>> 0, arg2);
+        },
+        __wbg_setPipeline_0c34cc40ab8d6499: function(arg0, arg1) {
+            arg0.setPipeline(arg1);
+        },
+        __wbg_set_62f340d5d135b4db: function(arg0, arg1, arg2) {
+            arg0.set(arg1, arg2 >>> 0);
+        },
+        __wbg_set_8ee2d34facb8466e: function() { return handleError(function (arg0, arg1, arg2) {
+            const ret = Reflect.set(arg0, arg1, arg2);
+            return ret;
+        }, arguments); },
+        __wbg_set_access_1cc7ab8607a9643c: function(arg0, arg1) {
+            arg0.access = __wbindgen_enum_GpuStorageTextureAccess[arg1];
+        },
+        __wbg_set_beginning_of_pass_write_index_ac45c363336c24c7: function(arg0, arg1) {
+            arg0.beginningOfPassWriteIndex = arg1 >>> 0;
+        },
+        __wbg_set_bind_group_layouts_b4667372bdcee99f: function(arg0, arg1) {
+            arg0.bindGroupLayouts = arg1;
+        },
+        __wbg_set_binding_0a48264269982c5e: function(arg0, arg1) {
+            arg0.binding = arg1 >>> 0;
+        },
+        __wbg_set_binding_15ab1e2c74990b25: function(arg0, arg1) {
+            arg0.binding = arg1 >>> 0;
+        },
+        __wbg_set_buffer_3b3e4c4a884d1610: function(arg0, arg1) {
+            arg0.buffer = arg1;
+        },
+        __wbg_set_buffer_ff433f6fc0bcc260: function(arg0, arg1) {
+            arg0.buffer = arg1;
+        },
+        __wbg_set_code_c616b86ce504e24a: function(arg0, arg1, arg2) {
+            arg0.code = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_compute_7c274f1347709d07: function(arg0, arg1) {
+            arg0.compute = arg1;
+        },
+        __wbg_set_create_0654e513e8ccb2be: function(arg0, arg1) {
+            arg0.create = arg1 !== 0;
+        },
+        __wbg_set_create_4b5cddb7e7c14744: function(arg0, arg1) {
+            arg0.create = arg1 !== 0;
+        },
+        __wbg_set_end_of_pass_write_index_c60088bc589e6882: function(arg0, arg1) {
+            arg0.endOfPassWriteIndex = arg1 >>> 0;
+        },
+        __wbg_set_entries_bfc700c1f97eec0b: function(arg0, arg1) {
+            arg0.entries = arg1;
+        },
+        __wbg_set_entries_f07df780e3613292: function(arg0, arg1) {
+            arg0.entries = arg1;
+        },
+        __wbg_set_entry_point_aa503b3bb9fed987: function(arg0, arg1, arg2) {
+            arg0.entryPoint = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_format_b8158198b657d617: function(arg0, arg1) {
+            arg0.format = __wbindgen_enum_GpuTextureFormat[arg1];
+        },
+        __wbg_set_has_dynamic_offset_4d5601049080763e: function(arg0, arg1) {
+            arg0.hasDynamicOffset = arg1 !== 0;
+        },
+        __wbg_set_label_392dc66ad76d942d: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_3e06143ad04772ae: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_50f397060b5b5610: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_68e2953cfd33a5a5: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_76c4f74a38ff9bcd: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_79484ec4d6d85bbf: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_861c8e348e26599d: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_d1b6a326332d0520: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_d687cfb9a30329c8: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_label_e345704005fb385b: function(arg0, arg1, arg2) {
+            arg0.label = getStringFromWasm0(arg1, arg2);
+        },
+        __wbg_set_layout_b9b36c291ee7f2e1: function(arg0, arg1) {
+            arg0.layout = arg1;
+        },
+        __wbg_set_layout_cccbb8f794df887c: function(arg0, arg1) {
+            arg0.layout = arg1;
+        },
+        __wbg_set_mapped_at_creation_34da9d6bf64b78d6: function(arg0, arg1) {
+            arg0.mappedAtCreation = arg1 !== 0;
+        },
+        __wbg_set_min_binding_size_9389ad67218af140: function(arg0, arg1) {
+            arg0.minBindingSize = arg1;
+        },
+        __wbg_set_module_5f33a55198ad797f: function(arg0, arg1) {
+            arg0.module = arg1;
+        },
+        __wbg_set_multisampled_b526741755338725: function(arg0, arg1) {
+            arg0.multisampled = arg1 !== 0;
+        },
+        __wbg_set_offset_1a0f95ffb7dd6f40: function(arg0, arg1) {
+            arg0.offset = arg1;
+        },
+        __wbg_set_power_preference_915480f4b9565dc2: function(arg0, arg1) {
+            arg0.powerPreference = __wbindgen_enum_GpuPowerPreference[arg1];
+        },
+        __wbg_set_query_set_0a78c3dcb3650b2b: function(arg0, arg1) {
+            arg0.querySet = arg1;
+        },
+        __wbg_set_required_features_42347bf311233eb6: function(arg0, arg1) {
+            arg0.requiredFeatures = arg1;
+        },
+        __wbg_set_resource_f2d72f59cc9308fc: function(arg0, arg1) {
+            arg0.resource = arg1;
+        },
+        __wbg_set_sample_type_6d1e240a417bdf44: function(arg0, arg1) {
+            arg0.sampleType = __wbindgen_enum_GpuTextureSampleType[arg1];
+        },
+        __wbg_set_sampler_f864a162bad4f66f: function(arg0, arg1) {
+            arg0.sampler = arg1;
+        },
+        __wbg_set_size_6b2fc4a0e39e4d07: function(arg0, arg1) {
+            arg0.size = arg1;
+        },
+        __wbg_set_size_c78ae8d2e2181815: function(arg0, arg1) {
+            arg0.size = arg1;
+        },
+        __wbg_set_storage_texture_c3919f22b211c542: function(arg0, arg1) {
+            arg0.storageTexture = arg1;
+        },
+        __wbg_set_texture_bf820de044f0d291: function(arg0, arg1) {
+            arg0.texture = arg1;
+        },
+        __wbg_set_timestamp_writes_b9e1d87e2f057bd1: function(arg0, arg1) {
+            arg0.timestampWrites = arg1;
+        },
+        __wbg_set_type_40f4ae4fa32946cd: function(arg0, arg1) {
+            arg0.type = __wbindgen_enum_GpuBufferBindingType[arg1];
+        },
+        __wbg_set_type_4f1cd48d79f4d6dc: function(arg0, arg1) {
+            arg0.type = __wbindgen_enum_GpuSamplerBindingType[arg1];
+        },
+        __wbg_set_usage_9aa23fa1e13799a8: function(arg0, arg1) {
+            arg0.usage = arg1 >>> 0;
+        },
+        __wbg_set_view_dimension_36c0bf530395d014: function(arg0, arg1) {
+            arg0.viewDimension = __wbindgen_enum_GpuTextureViewDimension[arg1];
+        },
+        __wbg_set_view_dimension_553cd9fa176d06ca: function(arg0, arg1) {
+            arg0.viewDimension = __wbindgen_enum_GpuTextureViewDimension[arg1];
+        },
+        __wbg_set_visibility_eef2d8e9608a8981: function(arg0, arg1) {
+            arg0.visibility = arg1 >>> 0;
+        },
+        __wbg_size_7306c9406e13bf29: function(arg0) {
+            const ret = arg0.size;
+            return ret;
+        },
+        __wbg_stack_3b0d974bbf31e44f: function(arg0, arg1) {
+            const ret = arg1.stack;
+            const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        },
+        __wbg_static_accessor_GLOBAL_THIS_a1248013d790bf5f: function() {
+            const ret = typeof globalThis === 'undefined' ? null : globalThis;
+            return isLikeNone(ret) ? 0 : addToExternrefTable0(ret);
+        },
+        __wbg_static_accessor_GLOBAL_f2e0f995a21329ff: function() {
+            const ret = typeof global === 'undefined' ? null : global;
+            return isLikeNone(ret) ? 0 : addToExternrefTable0(ret);
+        },
+        __wbg_static_accessor_SELF_24f78b6d23f286ea: function() {
+            const ret = typeof self === 'undefined' ? null : self;
+            return isLikeNone(ret) ? 0 : addToExternrefTable0(ret);
+        },
+        __wbg_static_accessor_WINDOW_59fd959c540fe405: function() {
+            const ret = typeof window === 'undefined' ? null : window;
+            return isLikeNone(ret) ? 0 : addToExternrefTable0(ret);
+        },
+        __wbg_status_44ecb0ac1da253f4: function(arg0) {
+            const ret = arg0.status;
+            return ret;
+        },
+        __wbg_storage_8f8e63186ec77353: function(arg0) {
+            const ret = arg0.storage;
+            return ret;
+        },
+        __wbg_submit_2521bdd9a232bca7: function(arg0, arg1) {
+            arg0.submit(arg1);
+        },
+        __wbg_then_00eed3ac0b8e82cb: function(arg0, arg1, arg2) {
+            const ret = arg0.then(arg1, arg2);
+            return ret;
+        },
+        __wbg_then_479d77cb064907ee: function(arg0, arg1, arg2) {
+            const ret = arg0.then(arg1, arg2);
+            return ret;
+        },
+        __wbg_then_a0c8db0381c8994c: function(arg0, arg1) {
+            const ret = arg0.then(arg1);
+            return ret;
+        },
+        __wbg_unmap_815a075fd850cb73: function(arg0) {
+            arg0.unmap();
+        },
+        __wbg_userAgent_d58193cc32293b16: function() { return handleError(function (arg0, arg1) {
+            const ret = arg1.userAgent;
+            const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+            const len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        }, arguments); },
+        __wbg_writeBuffer_e8b792fb0962f30d: function() { return handleError(function (arg0, arg1, arg2, arg3, arg4, arg5) {
+            arg0.writeBuffer(arg1, arg2, arg3, arg4, arg5);
+        }, arguments); },
+        __wbg_write_fc53b37dcc29642e: function() { return handleError(function (arg0, arg1, arg2) {
+            const ret = arg0.write(getArrayU8FromWasm0(arg1, arg2));
+            return ret;
+        }, arguments); },
+        __wbindgen_cast_0000000000000001: function(arg0, arg1) {
+            // Cast intrinsic for `Closure(Closure { owned: true, function: Function { arguments: [Externref], shim_idx: 155, ret: Unit, inner_ret: Some(Unit) }, mutable: true }) -> Externref`.
+            const ret = makeMutClosure(arg0, arg1, wasm_bindgen__convert__closures_____invoke__h235e00bf230ad8a4);
+            return ret;
+        },
+        __wbindgen_cast_0000000000000002: function(arg0, arg1) {
+            // Cast intrinsic for `Closure(Closure { owned: true, function: Function { arguments: [Externref], shim_idx: 177, ret: Result(Unit), inner_ret: Some(Result(Unit)) }, mutable: true }) -> Externref`.
+            const ret = makeMutClosure(arg0, arg1, wasm_bindgen__convert__closures_____invoke__h7ed8ea06cc0c8ca5);
+            return ret;
+        },
+        __wbindgen_cast_0000000000000003: function(arg0) {
+            // Cast intrinsic for `F64 -> Externref`.
+            const ret = arg0;
+            return ret;
+        },
+        __wbindgen_cast_0000000000000004: function(arg0, arg1) {
+            // Cast intrinsic for `Ref(Slice(U8)) -> NamedExternref("Uint8Array")`.
+            const ret = getArrayU8FromWasm0(arg0, arg1);
+            return ret;
+        },
+        __wbindgen_cast_0000000000000005: function(arg0, arg1) {
+            // Cast intrinsic for `Ref(String) -> Externref`.
+            const ret = getStringFromWasm0(arg0, arg1);
+            return ret;
+        },
+        __wbindgen_init_externref_table: function() {
+            const table = wasm.__wbindgen_externrefs;
+            const offset = table.grow(4);
+            table.set(0, undefined);
+            table.set(offset + 0, undefined);
+            table.set(offset + 1, null);
+            table.set(offset + 2, true);
+            table.set(offset + 3, false);
+        },
+    };
+    return {
+        __proto__: null,
+        "./flare_web_bg.js": import0,
+    };
+}
+
+function wasm_bindgen__convert__closures_____invoke__h235e00bf230ad8a4(arg0, arg1, arg2) {
+    wasm.wasm_bindgen__convert__closures_____invoke__h235e00bf230ad8a4(arg0, arg1, arg2);
+}
+
+function wasm_bindgen__convert__closures_____invoke__h7ed8ea06cc0c8ca5(arg0, arg1, arg2) {
+    const ret = wasm.wasm_bindgen__convert__closures_____invoke__h7ed8ea06cc0c8ca5(arg0, arg1, arg2);
+    if (ret[1]) {
+        throw takeFromExternrefTable0(ret[0]);
+    }
+}
+
+function wasm_bindgen__convert__closures_____invoke__hcdfd434894ba1863(arg0, arg1, arg2, arg3) {
+    wasm.wasm_bindgen__convert__closures_____invoke__hcdfd434894ba1863(arg0, arg1, arg2, arg3);
+}
+
+
+const __wbindgen_enum_GpuBufferBindingType = ["uniform", "storage", "read-only-storage"];
+
+
+const __wbindgen_enum_GpuPowerPreference = ["low-power", "high-performance"];
+
+
+const __wbindgen_enum_GpuSamplerBindingType = ["filtering", "non-filtering", "comparison"];
+
+
+const __wbindgen_enum_GpuStorageTextureAccess = ["write-only", "read-only", "read-write"];
+
+
+const __wbindgen_enum_GpuTextureFormat = ["r8unorm", "r8snorm", "r8uint", "r8sint", "r16uint", "r16sint", "r16float", "rg8unorm", "rg8snorm", "rg8uint", "rg8sint", "r32uint", "r32sint", "r32float", "rg16uint", "rg16sint", "rg16float", "rgba8unorm", "rgba8unorm-srgb", "rgba8snorm", "rgba8uint", "rgba8sint", "bgra8unorm", "bgra8unorm-srgb", "rgb9e5ufloat", "rgb10a2uint", "rgb10a2unorm", "rg11b10ufloat", "rg32uint", "rg32sint", "rg32float", "rgba16uint", "rgba16sint", "rgba16float", "rgba32uint", "rgba32sint", "rgba32float", "stencil8", "depth16unorm", "depth24plus", "depth24plus-stencil8", "depth32float", "depth32float-stencil8", "bc1-rgba-unorm", "bc1-rgba-unorm-srgb", "bc2-rgba-unorm", "bc2-rgba-unorm-srgb", "bc3-rgba-unorm", "bc3-rgba-unorm-srgb", "bc4-r-unorm", "bc4-r-snorm", "bc5-rg-unorm", "bc5-rg-snorm", "bc6h-rgb-ufloat", "bc6h-rgb-float", "bc7-rgba-unorm", "bc7-rgba-unorm-srgb", "etc2-rgb8unorm", "etc2-rgb8unorm-srgb", "etc2-rgb8a1unorm", "etc2-rgb8a1unorm-srgb", "etc2-rgba8unorm", "etc2-rgba8unorm-srgb", "eac-r11unorm", "eac-r11snorm", "eac-rg11unorm", "eac-rg11snorm", "astc-4x4-unorm", "astc-4x4-unorm-srgb", "astc-5x4-unorm", "astc-5x4-unorm-srgb", "astc-5x5-unorm", "astc-5x5-unorm-srgb", "astc-6x5-unorm", "astc-6x5-unorm-srgb", "astc-6x6-unorm", "astc-6x6-unorm-srgb", "astc-8x5-unorm", "astc-8x5-unorm-srgb", "astc-8x6-unorm", "astc-8x6-unorm-srgb", "astc-8x8-unorm", "astc-8x8-unorm-srgb", "astc-10x5-unorm", "astc-10x5-unorm-srgb", "astc-10x6-unorm", "astc-10x6-unorm-srgb", "astc-10x8-unorm", "astc-10x8-unorm-srgb", "astc-10x10-unorm", "astc-10x10-unorm-srgb", "astc-12x10-unorm", "astc-12x10-unorm-srgb", "astc-12x12-unorm", "astc-12x12-unorm-srgb"];
+
+
+const __wbindgen_enum_GpuTextureSampleType = ["float", "unfilterable-float", "depth", "sint", "uint"];
+
+
+const __wbindgen_enum_GpuTextureViewDimension = ["1d", "2d", "2d-array", "cube", "cube-array", "3d"];
+const FlareEngineFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_flareengine_free(ptr >>> 0, 1));
+const FlareProgressiveLoaderFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_flareprogressiveloader_free(ptr >>> 0, 1));
+const FlareTokenizerFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_flaretokenizer_free(ptr >>> 0, 1));
+
+function addToExternrefTable0(obj) {
+    const idx = wasm.__externref_table_alloc();
+    wasm.__wbindgen_externrefs.set(idx, obj);
+    return idx;
+}
+
+const CLOSURE_DTORS = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(state => wasm.__wbindgen_destroy_closure(state.a, state.b));
+
+function debugString(val) {
+    // primitive types
+    const type = typeof val;
+    if (type == 'number' || type == 'boolean' || val == null) {
+        return  `${val}`;
+    }
+    if (type == 'string') {
+        return `"${val}"`;
+    }
+    if (type == 'symbol') {
+        const description = val.description;
+        if (description == null) {
+            return 'Symbol';
+        } else {
+            return `Symbol(${description})`;
+        }
+    }
+    if (type == 'function') {
+        const name = val.name;
+        if (typeof name == 'string' && name.length > 0) {
+            return `Function(${name})`;
+        } else {
+            return 'Function';
+        }
+    }
+    // objects
+    if (Array.isArray(val)) {
+        const length = val.length;
+        let debug = '[';
+        if (length > 0) {
+            debug += debugString(val[0]);
+        }
+        for(let i = 1; i < length; i++) {
+            debug += ', ' + debugString(val[i]);
+        }
+        debug += ']';
+        return debug;
+    }
+    // Test for built-in
+    const builtInMatches = /\[object ([^\]]+)\]/.exec(toString.call(val));
+    let className;
+    if (builtInMatches && builtInMatches.length > 1) {
+        className = builtInMatches[1];
+    } else {
+        // Failed to match the standard '[object ClassName]'
+        return toString.call(val);
+    }
+    if (className == 'Object') {
+        // we're a user defined class or Object
+        // JSON.stringify avoids problems with cycles, and is generally much
+        // easier than looping through ownProperties of `val`.
+        try {
+            return 'Object(' + JSON.stringify(val) + ')';
+        } catch (_) {
+            return 'Object';
+        }
+    }
+    // errors
+    if (val instanceof Error) {
+        return `${val.name}: ${val.message}\n${val.stack}`;
+    }
+    // TODO we could test for more things here, like `Set`s and `Map`s.
+    return className;
+}
+
+function getArrayF32FromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return getFloat32ArrayMemory0().subarray(ptr / 4, ptr / 4 + len);
+}
+
+function getArrayU32FromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return getUint32ArrayMemory0().subarray(ptr / 4, ptr / 4 + len);
+}
+
+function getArrayU8FromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return getUint8ArrayMemory0().subarray(ptr / 1, ptr / 1 + len);
+}
+
+let cachedDataViewMemory0 = null;
+function getDataViewMemory0() {
+    if (cachedDataViewMemory0 === null || cachedDataViewMemory0.buffer.detached === true || (cachedDataViewMemory0.buffer.detached === undefined && cachedDataViewMemory0.buffer !== wasm.memory.buffer)) {
+        cachedDataViewMemory0 = new DataView(wasm.memory.buffer);
+    }
+    return cachedDataViewMemory0;
+}
+
+let cachedFloat32ArrayMemory0 = null;
+function getFloat32ArrayMemory0() {
+    if (cachedFloat32ArrayMemory0 === null || cachedFloat32ArrayMemory0.byteLength === 0) {
+        cachedFloat32ArrayMemory0 = new Float32Array(wasm.memory.buffer);
+    }
+    return cachedFloat32ArrayMemory0;
+}
+
+function getStringFromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return decodeText(ptr, len);
+}
+
+let cachedUint32ArrayMemory0 = null;
+function getUint32ArrayMemory0() {
+    if (cachedUint32ArrayMemory0 === null || cachedUint32ArrayMemory0.byteLength === 0) {
+        cachedUint32ArrayMemory0 = new Uint32Array(wasm.memory.buffer);
+    }
+    return cachedUint32ArrayMemory0;
+}
+
+let cachedUint8ArrayMemory0 = null;
+function getUint8ArrayMemory0() {
+    if (cachedUint8ArrayMemory0 === null || cachedUint8ArrayMemory0.byteLength === 0) {
+        cachedUint8ArrayMemory0 = new Uint8Array(wasm.memory.buffer);
+    }
+    return cachedUint8ArrayMemory0;
+}
+
+function handleError(f, args) {
+    try {
+        return f.apply(this, args);
+    } catch (e) {
+        const idx = addToExternrefTable0(e);
+        wasm.__wbindgen_exn_store(idx);
+    }
+}
+
+function isLikeNone(x) {
+    return x === undefined || x === null;
+}
+
+function makeMutClosure(arg0, arg1, f) {
+    const state = { a: arg0, b: arg1, cnt: 1 };
+    const real = (...args) => {
+
+        // First up with a closure we increment the internal reference
+        // count. This ensures that the Rust closure environment won't
+        // be deallocated while we're invoking it.
+        state.cnt++;
+        const a = state.a;
+        state.a = 0;
+        try {
+            return f(a, state.b, ...args);
+        } finally {
+            state.a = a;
+            real._wbg_cb_unref();
+        }
+    };
+    real._wbg_cb_unref = () => {
+        if (--state.cnt === 0) {
+            wasm.__wbindgen_destroy_closure(state.a, state.b);
+            state.a = 0;
+            CLOSURE_DTORS.unregister(state);
+        }
+    };
+    CLOSURE_DTORS.register(real, state, state);
+    return real;
+}
+
+function passArray32ToWasm0(arg, malloc) {
+    const ptr = malloc(arg.length * 4, 4) >>> 0;
+    getUint32ArrayMemory0().set(arg, ptr / 4);
+    WASM_VECTOR_LEN = arg.length;
+    return ptr;
+}
+
+function passArray8ToWasm0(arg, malloc) {
+    const ptr = malloc(arg.length * 1, 1) >>> 0;
+    getUint8ArrayMemory0().set(arg, ptr / 1);
+    WASM_VECTOR_LEN = arg.length;
+    return ptr;
+}
+
+function passArrayF32ToWasm0(arg, malloc) {
+    const ptr = malloc(arg.length * 4, 4) >>> 0;
+    getFloat32ArrayMemory0().set(arg, ptr / 4);
+    WASM_VECTOR_LEN = arg.length;
+    return ptr;
+}
+
+function passStringToWasm0(arg, malloc, realloc) {
+    if (realloc === undefined) {
+        const buf = cachedTextEncoder.encode(arg);
+        const ptr = malloc(buf.length, 1) >>> 0;
+        getUint8ArrayMemory0().subarray(ptr, ptr + buf.length).set(buf);
+        WASM_VECTOR_LEN = buf.length;
+        return ptr;
+    }
+
+    let len = arg.length;
+    let ptr = malloc(len, 1) >>> 0;
+
+    const mem = getUint8ArrayMemory0();
+
+    let offset = 0;
+
+    for (; offset < len; offset++) {
+        const code = arg.charCodeAt(offset);
+        if (code > 0x7F) break;
+        mem[ptr + offset] = code;
+    }
+    if (offset !== len) {
+        if (offset !== 0) {
+            arg = arg.slice(offset);
+        }
+        ptr = realloc(ptr, len, len = offset + arg.length * 3, 1) >>> 0;
+        const view = getUint8ArrayMemory0().subarray(ptr + offset, ptr + len);
+        const ret = cachedTextEncoder.encodeInto(arg, view);
+
+        offset += ret.written;
+        ptr = realloc(ptr, len, offset, 1) >>> 0;
+    }
+
+    WASM_VECTOR_LEN = offset;
+    return ptr;
+}
+
+function takeFromExternrefTable0(idx) {
+    const value = wasm.__wbindgen_externrefs.get(idx);
+    wasm.__externref_table_dealloc(idx);
+    return value;
+}
+
+let cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+cachedTextDecoder.decode();
+const MAX_SAFARI_DECODE_BYTES = 2146435072;
+let numBytesDecoded = 0;
+function decodeText(ptr, len) {
+    numBytesDecoded += len;
+    if (numBytesDecoded >= MAX_SAFARI_DECODE_BYTES) {
+        cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+        cachedTextDecoder.decode();
+        numBytesDecoded = len;
+    }
+    return cachedTextDecoder.decode(getUint8ArrayMemory0().subarray(ptr, ptr + len));
+}
+
+const cachedTextEncoder = new TextEncoder();
+
+if (!('encodeInto' in cachedTextEncoder)) {
+    cachedTextEncoder.encodeInto = function (arg, view) {
+        const buf = cachedTextEncoder.encode(arg);
+        view.set(buf);
+        return {
+            read: arg.length,
+            written: buf.length
+        };
+    };
+}
+
+let WASM_VECTOR_LEN = 0;
+
+let wasmModule, wasm;
+function __wbg_finalize_init(instance, module) {
+    wasm = instance.exports;
+    wasmModule = module;
+    cachedDataViewMemory0 = null;
+    cachedFloat32ArrayMemory0 = null;
+    cachedUint32ArrayMemory0 = null;
+    cachedUint8ArrayMemory0 = null;
+    wasm.__wbindgen_start();
+    return wasm;
+}
+
+async function __wbg_load(module, imports) {
+    if (typeof Response === 'function' && module instanceof Response) {
+        if (typeof WebAssembly.instantiateStreaming === 'function') {
+            try {
+                return await WebAssembly.instantiateStreaming(module, imports);
+            } catch (e) {
+                const validResponse = module.ok && expectedResponseType(module.type);
+
+                if (validResponse && module.headers.get('Content-Type') !== 'application/wasm') {
+                    console.warn("`WebAssembly.instantiateStreaming` failed because your server does not serve Wasm with `application/wasm` MIME type. Falling back to `WebAssembly.instantiate` which is slower. Original error:\n", e);
+
+                } else { throw e; }
+            }
+        }
+
+        const bytes = await module.arrayBuffer();
+        return await WebAssembly.instantiate(bytes, imports);
+    } else {
+        const instance = await WebAssembly.instantiate(module, imports);
+
+        if (instance instanceof WebAssembly.Instance) {
+            return { instance, module };
+        } else {
+            return instance;
+        }
+    }
+
+    function expectedResponseType(type) {
+        switch (type) {
+            case 'basic': case 'cors': case 'default': return true;
+        }
+        return false;
+    }
+}
+
+function initSync(module) {
+    if (wasm !== undefined) return wasm;
+
+
+    if (module !== undefined) {
+        if (Object.getPrototypeOf(module) === Object.prototype) {
+            ({module} = module)
+        } else {
+            console.warn('using deprecated parameters for `initSync()`; pass a single object instead')
+        }
+    }
+
+    const imports = __wbg_get_imports();
+    if (!(module instanceof WebAssembly.Module)) {
+        module = new WebAssembly.Module(module);
+    }
+    const instance = new WebAssembly.Instance(module, imports);
+    return __wbg_finalize_init(instance, module);
+}
+
+async function __wbg_init(module_or_path) {
+    if (wasm !== undefined) return wasm;
+
+
+    if (module_or_path !== undefined) {
+        if (Object.getPrototypeOf(module_or_path) === Object.prototype) {
+            ({module_or_path} = module_or_path)
+        } else {
+            console.warn('using deprecated parameters for the initialization function; pass a single object instead')
+        }
+    }
+
+    if (module_or_path === undefined) {
+        module_or_path = new URL('flare_web_bg.wasm', import.meta.url);
+    }
+    const imports = __wbg_get_imports();
+
+    if (typeof module_or_path === 'string' || (typeof Request === 'function' && module_or_path instanceof Request) || (typeof URL === 'function' && module_or_path instanceof URL)) {
+        module_or_path = fetch(module_or_path);
+    }
+
+    const { instance, module } = await __wbg_load(await module_or_path, imports);
+
+    return __wbg_finalize_init(instance, module);
+}
+
+export { initSync, __wbg_init as default };