From fa95056341a0d03270abc656944586091c33ce25 Mon Sep 17 00:00:00 2001
From: danielhanchen <danielhanchen@gmail.com>
Date: Sat, 8 Aug 2026 05:30:23 +0000
Subject: [PATCH] CI: let a cold CUDA profile warm its ccache from a sibling
 profile

---
 .github/workflows/unsloth-prebuilt-cuda-windows.yml | 8 ++++++++
 .github/workflows/unsloth-prebuilt-cuda.yml         | 8 ++++++++
 2 files changed, 16 insertions(+)

diff --git a/.github/workflows/unsloth-prebuilt-cuda-windows.yml b/.github/workflows/unsloth-prebuilt-cuda-windows.yml
index 16772e29d59c..82fdcad0d110 100644
--- a/.github/workflows/unsloth-prebuilt-cuda-windows.yml
+++ b/.github/workflows/unsloth-prebuilt-cuda-windows.yml
@@ -118,8 +118,16 @@ jobs:
         uses: hendrikmuhs/ccache-action@v1.2.23
         with:
           key: cuda-${{ matrix.cuda }}-windows-${{ matrix.profile }}-${{ inputs.tag }}
+          # The second, broader key is a fallback: when this profile has no
+          # cache of its own (first build of a new profile, or its last save was
+          # lost) it warms from a sibling profile of the same CUDA line and OS
+          # instead of starting cold. Only the CUDA arch list differs between
+          # profiles, and that affects .cu compilation only, so every plain C++
+          # translation unit -- the large majority -- has a byte-identical
+          # command line and hits.
           restore-keys: |
             cuda-${{ matrix.cuda }}-windows-${{ matrix.profile }}
+            cuda-${{ matrix.cuda }}-windows-
           append-timestamp: false
           variant: ccache
           max-size: 2G
diff --git a/.github/workflows/unsloth-prebuilt-cuda.yml b/.github/workflows/unsloth-prebuilt-cuda.yml
index 29d188aeea21..342e5d62d7b4 100644
--- a/.github/workflows/unsloth-prebuilt-cuda.yml
+++ b/.github/workflows/unsloth-prebuilt-cuda.yml
@@ -96,8 +96,16 @@ jobs:
         uses: hendrikmuhs/ccache-action@v1.2.23
         with:
           key: cuda-${{ matrix.cuda }}-${{ matrix.arch }}-${{ matrix.profile }}-${{ inputs.tag }}
+          # The second, broader key is a fallback: when this profile has no
+          # cache of its own (first build of a new profile, or its last save was
+          # lost) it warms from a sibling profile of the same CUDA line and OS
+          # instead of starting cold. Only the CUDA arch list differs between
+          # profiles, and that affects .cu compilation only, so every plain C++
+          # translation unit -- the large majority -- has a byte-identical
+          # command line and hits.
           restore-keys: |
             cuda-${{ matrix.cuda }}-${{ matrix.arch }}-${{ matrix.profile }}
+            cuda-${{ matrix.cuda }}-${{ matrix.arch }}-
           append-timestamp: false
           variant: ccache
           max-size: 2G