@pie-project/server-web 0.6.0 → 0.6.1

package/dist/pie_browser_bg.wasm

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/pie_browser_bg.wasm
-size: 34261602 bytes
-sha256: e3876081c93d93e61a359a741db0c63373bb77995bfca92c0071f99481a17e8d
+size: 34288832 bytes
+sha256: 1c2062bbd50c252fdc437b37d6cc863d62836216745bc7138c9b85ba49165ee1

package/dist/precompile.mjs

@@ -509,7 +509,7 @@ function __wbg_get_imports() {
           const a = state0.a;
           state0.a = 0;
           try {
-            return __wasm_bindgen_func_elem_55767(a, state0.b, arg02, arg12);
+            return __wasm_bindgen_func_elem_55690(a, state0.b, arg02, arg12);
           } finally {
             state0.a = a;
           }
@@ -850,23 +850,23 @@ function __wbg_get_imports() {
       }, arguments);
     },
     __wbindgen_generic_0000000000000001: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_55765);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_55688);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000002: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11488);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11414);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000003: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11489);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11415);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000004: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11488_15);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11414_15);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000005: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11488_16);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11414_16);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000006: function(arg0) {
@@ -898,16 +898,16 @@ function __wbg_get_imports() {
     "./platform.mjs": platform_exports
   };
 }
-function __wasm_bindgen_func_elem_11489(arg0, arg1, arg2) {
-  wasm2.__wasm_bindgen_func_elem_11489(arg0, arg1, addHeapObject(arg2));
+function __wasm_bindgen_func_elem_11415(arg0, arg1, arg2) {
+  wasm2.__wasm_bindgen_func_elem_11415(arg0, arg1, addHeapObject(arg2));
 }
-function __wasm_bindgen_func_elem_55767(arg0, arg1, arg2, arg3) {
-  wasm2.__wasm_bindgen_func_elem_55767(arg0, arg1, addHeapObject(arg2), addHeapObject(arg3));
+function __wasm_bindgen_func_elem_55690(arg0, arg1, arg2, arg3) {
+  wasm2.__wasm_bindgen_func_elem_55690(arg0, arg1, addHeapObject(arg2), addHeapObject(arg3));
 }
-function __wasm_bindgen_func_elem_55765(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_55688(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_55765(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_55688(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {
@@ -917,10 +917,10 @@ function __wasm_bindgen_func_elem_55765(arg0, arg1, arg2) {
     wasm2.__wbindgen_add_to_stack_pointer(16);
   }
 }
-function __wasm_bindgen_func_elem_11488(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_11414(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_11488(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_11414(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {
@@ -930,10 +930,10 @@ function __wasm_bindgen_func_elem_11488(arg0, arg1, arg2) {
     wasm2.__wbindgen_add_to_stack_pointer(16);
   }
 }
-function __wasm_bindgen_func_elem_11488_15(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_11414_15(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_11488_15(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_11414_15(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {
@@ -943,10 +943,10 @@ function __wasm_bindgen_func_elem_11488_15(arg0, arg1, arg2) {
     wasm2.__wbindgen_add_to_stack_pointer(16);
   }
 }
-function __wasm_bindgen_func_elem_11488_16(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_11414_16(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_11488_16(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_11414_16(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {

package/dist/worker.mjs

@@ -561,7 +561,7 @@ function __wbg_get_imports() {
           const a = state0.a;
           state0.a = 0;
           try {
-            return __wasm_bindgen_func_elem_55767(a, state0.b, arg02, arg12);
+            return __wasm_bindgen_func_elem_55690(a, state0.b, arg02, arg12);
           } finally {
             state0.a = a;
           }
@@ -902,23 +902,23 @@ function __wbg_get_imports() {
       }, arguments);
     },
     __wbindgen_generic_0000000000000001: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_55765);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_55688);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000002: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11488);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11414);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000003: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11489);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11415);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000004: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11488_15);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11414_15);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000005: function(arg0, arg1) {
-      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11488_16);
+      const ret = makeMutClosure(arg0, arg1, __wasm_bindgen_func_elem_11414_16);
       return addHeapObject(ret);
     },
     __wbindgen_generic_0000000000000006: function(arg0) {
@@ -950,16 +950,16 @@ function __wbg_get_imports() {
     "./platform.mjs": platform_exports
   };
 }
-function __wasm_bindgen_func_elem_11489(arg0, arg1, arg2) {
-  wasm2.__wasm_bindgen_func_elem_11489(arg0, arg1, addHeapObject(arg2));
+function __wasm_bindgen_func_elem_11415(arg0, arg1, arg2) {
+  wasm2.__wasm_bindgen_func_elem_11415(arg0, arg1, addHeapObject(arg2));
 }
-function __wasm_bindgen_func_elem_55767(arg0, arg1, arg2, arg3) {
-  wasm2.__wasm_bindgen_func_elem_55767(arg0, arg1, addHeapObject(arg2), addHeapObject(arg3));
+function __wasm_bindgen_func_elem_55690(arg0, arg1, arg2, arg3) {
+  wasm2.__wasm_bindgen_func_elem_55690(arg0, arg1, addHeapObject(arg2), addHeapObject(arg3));
 }
-function __wasm_bindgen_func_elem_55765(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_55688(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_55765(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_55688(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {
@@ -969,10 +969,10 @@ function __wasm_bindgen_func_elem_55765(arg0, arg1, arg2) {
     wasm2.__wbindgen_add_to_stack_pointer(16);
   }
 }
-function __wasm_bindgen_func_elem_11488(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_11414(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_11488(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_11414(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {
@@ -982,10 +982,10 @@ function __wasm_bindgen_func_elem_11488(arg0, arg1, arg2) {
     wasm2.__wbindgen_add_to_stack_pointer(16);
   }
 }
-function __wasm_bindgen_func_elem_11488_15(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_11414_15(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_11488_15(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_11414_15(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {
@@ -995,10 +995,10 @@ function __wasm_bindgen_func_elem_11488_15(arg0, arg1, arg2) {
     wasm2.__wbindgen_add_to_stack_pointer(16);
   }
 }
-function __wasm_bindgen_func_elem_11488_16(arg0, arg1, arg2) {
+function __wasm_bindgen_func_elem_11414_16(arg0, arg1, arg2) {
   try {
     const retptr = wasm2.__wbindgen_add_to_stack_pointer(-16);
-    wasm2.__wasm_bindgen_func_elem_11488_16(retptr, arg0, arg1, addHeapObject(arg2));
+    wasm2.__wasm_bindgen_func_elem_11414_16(retptr, arg0, arg1, addHeapObject(arg2));
     var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
     var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
     if (r1) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pie-project/server-web",
-  "version": "0.6.0",
+  "version": "0.6.1",
   "description": "The browser half of @pie-project/server: the runtime and the WebGPU engine as wasm, reached through `import { Server } from \"@pie-project/server\"` under a bundler.",
   "type": "module",
   "main": "dist/pie.mjs",

package/README.md

@@ -1,102 +0,0 @@
-# @pie-project/server-web
-
-The browser half of `@pie-project/server`: the same runtime, WebGPU engine
-and inferlet sandbox as `pie serve`, compiled to `wasm32-unknown-unknown`.
-No server-side inference: the page fetches a `.wgpu.zt` artifact, the
-weights land in WebGPU buffers, and inferlets run under wasmtime's Pulley
-interpreter inside the same module. A bundler reaches it through
-`@pie-project/server`'s `browser` export condition; nothing imports it by
-name.
-
-## Use
-
-The package ships the wasm and a self-contained bundle (`dist/pie.mjs`,
-`dist/worker.mjs`, `dist/pie_browser_bg.wasm`); the client
-(`@pie-project/client`) is bundled in and re-exported. The worker and the
-wasm are fetched relative to the module, so serve them from your own origin
-(a bundler that understands `new URL(…, import.meta.url)` — Vite, webpack 5 —
-does this by itself; without one, copy `dist/` next to your page).
-
-```js
-import { Server } from "@pie-project/server";      // resolves here under a bundler
-import "@pie-project/language-python";
-
-const server = await Server.start({ model: "/models/qwen3.5-0.8b.wgpu.zt" });   // served with Range support
-const program = await server.install("/inferlets/beam_search.wasm");   // a URL, or bytes with a file name
-const client = await server.connect();              // a PieClient, as for `pie serve`
-const proc = await client.launchProcess(program, { prompt: "The capital of France is", max_tokens: 16 });
-for (;;) {
-  const { event, value } = await proc.recv();
-  if (event === "return" || event === "error") break;
-}
-await server.shutdown();
-```
-
-The page-level functions behind the class (`load`, `bootLazy`, `install`,
-`installLanguage`, `connect`, `awaitCache`, `memoryBytes`) are exported as
-well; `tests/browser` drives them directly.
-
-
-- The page must be cross-origin isolated (`Cross-Origin-Opener-Policy:
-  same-origin`, `Cross-Origin-Embedder-Policy: require-corp`), and the
-  browser needs WebGPU and JSPI: Chrome 137+, Firefox 153+ (Firefox on
-  Linux has no WebGPU, so the engine cannot boot there).
-- `bootLazy` never holds the artifact in the tab: the loader's reads are
-  served as 8 MiB-aligned `Range` requests, and after a boot from the
-  network the artifact is copied into the origin's private file system
-  (files up to about 2 GB) so the next visit reads it from there;
-  `awaitCache()` waits for that copy. Peak wasm memory is a few times the
-  largest tensor: 0.8B boots in 285 MiB, 9B in 4 bits in 1.5 GiB.
-- The engine cannot ask WebGPU how much device memory there is and assumes
-  8 GiB; a large model with the default pools is refused ("raise the device
-  weight budget") — pass `device_memory_mb = 20480` (or a smaller
-  `device_kv_budget`) in the boot config. `Settings` in
-  `crates/worker/src/embedded.rs` lists the keys, plus `device_memory_mb` and
-  `power_preference` for the WebGPU engine.
-- A tab compiles a script language's component (Python's is 37 MB) on the
-  first inferlet in that language, about 8 s, and again after every reload.
-  Precompile it once at build time with this same build,
-  `node node_modules/@pie-project/server-web/dist/precompile.mjs python.wasm
-  python.pulley.cwasm` (or `precompileComponent` from
-  `@pie-project/server-web/precompile`; no GPU needed), serve the result
-  beside the source, and pass it as
-  `installLanguage("python", source, { precompiled })`. It fits only the
-  `pie_browser_bg.wasm` it was made with; one that does not is ignored with
-  a warning and the source compiled instead.
-- Rust, JavaScript and Python inferlets run. Not available in a tab: MoE
-  host-tier expert streaming, Python inferlet snapshots, telemetry, and
-  outbound HTTP or files from inferlets (the WASI interfaces link, a call
-  fails).
-- `index.d.ts` carries the types.
-
-## Build
-
-```bash
-rustup target add wasm32-unknown-unknown
-cargo install wasm-bindgen-cli --version 0.2.128 --locked   # must match Cargo.lock
-./packages/javascript/server-web/build.sh          # release; `min` for the 23 MB shipping build, `dev` for debug
-```
-
-`build.sh` builds the crate in `rust/`, runs wasm-bindgen into `pkg/`, and bundles
-`src/` with esbuild into `dist/` (`npm run bundle`). The package's own
-dependencies are installed on the first run. Tests, benchmarks and CI gates
-live under `tests/browser` (see its README).
-
-## How it fits together
-
-| piece | what it does |
-|---|---|
-| `rust/` (crate `pie-browser`) | The wasm-bindgen host: boot, install, sessions speaking the MessagePack frames `pie serve`'s WebSocket carries, `pie_tick`. |
-| `crates/wasmtime-web` | wasmtime in a tab: heap-backed "virtual memory", TLS slots, fibers implemented by the page's JSPI glue (`src/platform.mjs`), and the WASI host for the inferlet's imports (cli, io, clocks, random; 0.2 linked by hand, the 0.3 clocks from `crates/inferlet/wit`). The 0.2 filesystem and http interfaces a JavaScript or Python guest runtime imports regardless are stubbed per component, so they link and only a call fails. |
-| `crates/web-std` | The executor the page ticks, timers over `performance.now()`, crossbeam-shaped channels, and green threads on fibers. The runtime's scheduler and engine lanes run on those unchanged; `crates/runtime`'s `rt` facade names it on wasm32. |
-| `crates/engine-wgpu` | The same engine on both hosts, written for the browser: no GPU wait blocks, every step lands by callback, guests run from the landing. What differs by host sits in `device/host.rs`. |
-| `src/worker.mjs` | The host: wasm init, the tick loop, sessions, the artifact served by byte range. Runs in a Web Worker so the page stays responsive through the inferlet's in-browser compile and every forward pass; `load({ worker: false })` runs it on the page for debugging. |
-| `src/pie.mjs` | The page-side API: the `Server` class over `load`, `bootLazy`, `install`, `connect`. |
-| `src/languages.mjs` | The language registry `@pie-project/server/languages` re-exports: what a language package calls on import. |
-| `src/platform.mjs` | Fibers (JSPI), the clock, the wake hook the wasm imports. |
-| `src/transport.mjs` | The client's in-page socket over the runtime's frame API. |
-
-The native wgpu server is the engine's test host rather than a production
-backend: it runs the browser's code path over Vulkan, Metal or DX12 with a
-poll thread standing in for the page's event loop, so what the browser suite
-exercises is exercised natively too, and the two produce identical tokens.