Spaces:
Sleeping
Sleeping
Taylor commited on
Commit ·
90bf42d
1
Parent(s): 5b5a680
fix: handle WASM OOM for LM head + show error details
Browse filesLM head matVec is 49152 x 960 = 188MB, which can exceed WASM
linear memory. Fall back to JS for matrices >100MB.
Also: read error body from Aether 500 responses to show
actual error message instead of generic 'Internal Server Error'.
- aether-server.mjs +19 -9
- app.py +7 -0
aether-server.mjs
CHANGED
|
@@ -61,15 +61,25 @@ async function loadSIMD() {
|
|
| 61 |
|
| 62 |
return {
|
| 63 |
matVec(matrix, vector, rows, cols) {
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
},
|
| 74 |
rmsNorm(x, weight, eps) {
|
| 75 |
const saved = wasm.getHeapPtr();
|
|
|
|
| 61 |
|
| 62 |
return {
|
| 63 |
matVec(matrix, vector, rows, cols) {
|
| 64 |
+
// Fall back to JS for huge matrices (LM head: 49152 x 960 = 188MB)
|
| 65 |
+
// that exceed WASM linear memory
|
| 66 |
+
if (matrix.byteLength > 100_000_000) {
|
| 67 |
+
return matVecJS(matrix, vector, rows, cols);
|
| 68 |
+
}
|
| 69 |
+
try {
|
| 70 |
+
const saved = wasm.getHeapPtr();
|
| 71 |
+
const mPtr = wasm.allocate(matrix.byteLength);
|
| 72 |
+
const vPtr = wasm.allocate(vector.byteLength);
|
| 73 |
+
const rPtr = wasm.allocate(rows * 4);
|
| 74 |
+
copyTo(mPtr, matrix); copyTo(vPtr, vector);
|
| 75 |
+
wasm.matVecSimdBatch4(mPtr, vPtr, rPtr, rows, cols);
|
| 76 |
+
const result = copyFrom(rPtr, rows);
|
| 77 |
+
wasm.resetHeap(saved);
|
| 78 |
+
return result;
|
| 79 |
+
} catch (e) {
|
| 80 |
+
// WASM OOM -- fall back to JS
|
| 81 |
+
return matVecJS(matrix, vector, rows, cols);
|
| 82 |
+
}
|
| 83 |
},
|
| 84 |
rmsNorm(x, weight, eps) {
|
| 85 |
const saved = wasm.getHeapPtr();
|
app.py
CHANGED
|
@@ -96,6 +96,13 @@ def gen_aether(prompt):
|
|
| 96 |
result["tokens"],
|
| 97 |
result["avgTokenMs"],
|
| 98 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
except Exception as e:
|
| 100 |
return f"[Aether error: {e}]", 0, 0, 0
|
| 101 |
|
|
|
|
| 96 |
result["tokens"],
|
| 97 |
result["avgTokenMs"],
|
| 98 |
)
|
| 99 |
+
except urllib.error.HTTPError as e:
|
| 100 |
+
body = e.read().decode() if e.fp else str(e)
|
| 101 |
+
try:
|
| 102 |
+
detail = json.loads(body).get("error", body[:200])
|
| 103 |
+
except Exception:
|
| 104 |
+
detail = body[:200]
|
| 105 |
+
return f"[Aether error: {detail}]", 0, 0, 0
|
| 106 |
except Exception as e:
|
| 107 |
return f"[Aether error: {e}]", 0, 0, 0
|
| 108 |
|