From 51baadad9dc534bfe4dda128b551624b83667db4 Mon Sep 17 00:00:00 2001
From: Nick Mitchell <nickm@us.ibm.com>
Date: Fri, 20 Feb 2026 08:35:58 -0500
Subject: [PATCH] fix(ci): use stable cache key for CUDA flash-attn builds

The previous cache key included hashFiles Cargo.lock, which changes
every release, causing the cache to always miss and the expensive
Post save step to always run (~20GB upload for no benefit).

- Remove registry cache step (JS actions run on host in container
  jobs, cannot access the container ~/.cargo)
- Use stable key based on features + compute cap so the cache
  actually hits across releases
- On exact match, actions/cache skips Post save entirely
- Cargo fingerprinting still recompiles changed deps regardless

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Nick Mitchell <nickm@us.ibm.com>
---
 .github/workflows/release-cli.yml | 27 ++++++++++-----------------
 1 file changed, 10 insertions(+), 17 deletions(-)

diff --git a/.github/workflows/release-cli.yml b/.github/workflows/release-cli.yml
index bd63388b..3cf398f7 100644
--- a/.github/workflows/release-cli.yml
+++ b/.github/workflows/release-cli.yml
@@ -448,29 +448,22 @@ jobs:
         # Skip flash-attn variants on PRs (only run on release)
         if: github.event_name == 'release' || matrix.platform.features == 'cuda'
 
-      - name: Cache Cargo registry
-        # Only use cache for flash-attn variants on releases (they take longer to build)
-        if: github.event_name == 'release' && matrix.platform.features != 'cuda'
-        uses: actions/cache@v5
-        with:
-          path: |
-            ~/.cargo/registry/index/
-            ~/.cargo/registry/cache/
-            ~/.cargo/git/db/
-          key: linux-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
-          restore-keys: |
-            linux-cargo-registry-
-
       - name: Cache Cargo build
-        # Only use cache for flash-attn variants on releases (they take longer to build)
+        # Cache target/ for flash-attn variants on releases (they take ~45min to build).
+        # Note: ~/.cargo/registry is NOT cached because this is a container job — JS actions
+        # (like actions/cache) run on the host, which can't access the container's ~/.cargo.
+        # The target/ dir works because it's in the workspace, which is host-mounted.
+        #
+        # Key is stable (no Cargo.lock hash) so the cache actually hits across releases.
+        # On exact match, actions/cache skips the expensive Post save step entirely.
+        # Cargo's fingerprinting handles staleness — it recompiles changed deps regardless.
         if: github.event_name == 'release' && matrix.platform.features != 'cuda'
         uses: actions/cache@v5
         with:
           path: target/
-          key: linux-${{ matrix.platform.arch }}-cargo-build-cuda-${{ matrix.platform.features }}-${{ hashFiles('**/Cargo.lock') }}
+          key: cuda-build-${{ matrix.platform.features }}-${{ matrix.platform.cuda_compute_cap }}
           restore-keys: |
-            linux-${{ matrix.platform.arch }}-cargo-build-cuda-${{ matrix.platform.features }}-
-            linux-${{ matrix.platform.arch }}-cargo-build-cuda-
+            cuda-build-${{ matrix.platform.features }}-
 
       - name: Install build dependencies
         # Skip flash-attn variants on PRs (only run on release)