crispy-recall 0.1.2 → 0.1.3

package/dist/embed-pending.js

@@ -4594,6 +4594,11 @@ async function startServer() {
     "-ub",
     "8192",
     // micro-batch (ubatch) — also defaults to 512, must be raised
+    // Flash Attention: the attention compute buffer is otherwise O(n²) in the
+    // 8192-token batch (~3.7 GB on GPU), which OOMs cards <6 GB and forces CPU.
+    // -fa makes it O(n) (~0.8 GB → ~2.9 GB total), so 4 GB cards run on GPU,
+    // with numerically identical embeddings (verified cosine 1.0 vs non-FA).
+    "-fa",
     // nomic-embed-text-v1.5 is trained at 2048 ctx and uses Dynamic NTK-aware
     // RoPE scaling to extend to 8192. Without these flags, newer llama.cpp
     // (b9253+) refuses inputs >2048 tokens, and older versions silently
@@ -4801,6 +4806,9 @@ async function embedViaProcess(texts, modelPath) {
       "array",
       "-c",
       "8192",
+      // Flash Attention — see startServer: O(n) attention buffer instead of
+      // O(n²), so the one-shot path also fits cards <6 GB. Identical embeddings.
+      "-fa",
       // Match the YaRN flags on the llama-server batch path — nomic-embed-text-v1.5
       // is trained at 2048 ctx and extends to 8192 via Dynamic NTK-aware RoPE.
       // Without these, newer llama.cpp (b9253+) refuses inputs >2048 tokens here

package/dist/recall.js

@@ -3268,6 +3268,11 @@ async function startServer() {
     "-ub",
     "8192",
     // micro-batch (ubatch) — also defaults to 512, must be raised
+    // Flash Attention: the attention compute buffer is otherwise O(n²) in the
+    // 8192-token batch (~3.7 GB on GPU), which OOMs cards <6 GB and forces CPU.
+    // -fa makes it O(n) (~0.8 GB → ~2.9 GB total), so 4 GB cards run on GPU,
+    // with numerically identical embeddings (verified cosine 1.0 vs non-FA).
+    "-fa",
     // nomic-embed-text-v1.5 is trained at 2048 ctx and uses Dynamic NTK-aware
     // RoPE scaling to extend to 8192. Without these flags, newer llama.cpp
     // (b9253+) refuses inputs >2048 tokens, and older versions silently
@@ -3475,6 +3480,9 @@ async function embedViaProcess(texts, modelPath) {
       "array",
       "-c",
       "8192",
+      // Flash Attention — see startServer: O(n) attention buffer instead of
+      // O(n²), so the one-shot path also fits cards <6 GB. Identical embeddings.
+      "-fa",
       // Match the YaRN flags on the llama-server batch path — nomic-embed-text-v1.5
       // is trained at 2048 ctx and extends to 8192 via Dynamic NTK-aware RoPE.
       // Without these, newer llama.cpp (b9253+) refuses inputs >2048 tokens here
@@ -15016,6 +15024,12 @@ async function defaultProbe(args) {
     "array",
     "-c",
     "8192",
+    // Probe with the SAME flags the real embed server uses (esp. -fa). Without
+    // Flash Attention the probe allocates the ~3.7 GB O(n²) attention buffer and
+    // OOMs on a 4 GB card — failing the probe and falling back to CPU even
+    // though the actual -fa server (~2.9 GB) would have fit. Probe must mirror
+    // runtime to gate GPU adoption correctly.
+    "-fa",
     "--rope-scaling",
     "yarn",
     "--rope-freq-scale",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "crispy-recall",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "description": "Local session transcript memory for Claude Code and Codex — search past sessions with FTS5 + semantic vectors.",
   "license": "MIT",
   "author": "Sylvester Wong",