Studio whisper: pair slim bundles on the ggml commit, not the full llama tag (#7381)

The slim whisper bundle is ggml-less and links the ggml runtime out of the
installed llama.cpp prebuilt, so each whisper release pins a paired llama tag.
The gate required an exact tag match, but llama fork tags are
b<upstream_build>-mix-<ggml_commit> and the build number tracks upstream llama
and fork PRs that live outside ggml. When llama republishes a newer build with
the same ggml commit (a frequent event), the installed llama advances past the
whisper pin and curated dictation goes unavailable until whisper is republished,
even though the ggml runtime is ABI-identical.

Key the pairing gate on the ggml commit after -mix- instead of the full tag, in
all three comparison sites (slim_pairing_for_artifact,
_slim_release_incompatibility, resolve_selection). requires_ggml_sonames stays
the real per-file ABI gate, and a genuine ggml skew still fails closed. Tags
without a -mix- marker fall back to exact matching.

studio/install_whisper_prebuilt.py

@@ -335,6 +335,31 @@ def artifacts_for_host(
 
 
 # ── Slim selection (paired with the installed llama.cpp ggml runtime) ──
+def _llama_ggml_commit(tag: str) -> str | None:
+    """The ggml commit a llama.cpp fork tag was built against. Fork tags are
+    "b<upstream_build>-mix-<ggml_commit>"; the ggml commit after "-mix-" fixes
+    the ggml ABI the slim whisper bundle links against, while the build number
+    only tracks upstream llama / fork PRs outside ggml. None when the tag has no
+    "-mix-" marker (then only an exact tag pairs)."""
+    marker = "-mix-"
+    idx = tag.rfind(marker)
+    end = idx + len(marker)
+    return tag[end:] if idx >= 0 and end < len(tag) else None
+
+
+def llama_runtime_pairs(installed_tag: str, required_tag: Any) -> bool:
+    """Whether an installed llama tag can back a slim bundle needing required_tag.
+    An exact tag always pairs; so does a shared ggml commit, since a newer llama
+    build with the same ggml ships an ABI-identical runtime. requires_ggml_sonames
+    stays the real per-file ABI gate."""
+    if not isinstance(required_tag, str):
+        return False
+    if installed_tag == required_tag:
+        return True
+    commit = _llama_ggml_commit(installed_tag)
+    return commit is not None and commit == _llama_ggml_commit(required_tag)
+
+
 def slim_pairing_for_artifact(
     artifact: dict[str, Any], host: HostInfo, backend: str
 ) -> tuple[Path, str] | None:
@@ -348,10 +373,10 @@ def slim_pairing_for_artifact(
         return None
     llama_bin_dir, llama_tag, _profile = runtime
     requires_tag = artifact.get("requires_llama_tag")
-    if not isinstance(requires_tag, str) or requires_tag != llama_tag:
+    if not llama_runtime_pairs(llama_tag, requires_tag):
         log(
             f"slim_selection: {asset} skipped: installed llama tag {llama_tag!r} "
-            f"!= required {requires_tag!r}"
+            f"does not pair with required {requires_tag!r}"
         )
         return None
     sonames = artifact.get("requires_ggml_sonames")
@@ -466,11 +491,10 @@ def _slim_release_incompatibility(manifest: dict[str, Any], host: HostInfo) -> s
         for artifact in os_compatible
         if isinstance(artifact.get("requires_llama_tag"), str)
     }
-    if required_tags and installed_tag not in required_tags:
+    if required_tags and not any(llama_runtime_pairs(installed_tag, tag) for tag in required_tags):
         required_tag = sorted(required_tags)[0]
         return (
-            f"slim bundle requires llama.cpp {required_tag}; "
-            f"installed llama.cpp is {installed_tag}"
+            f"slim bundle requires llama.cpp {required_tag}; installed llama.cpp is {installed_tag}"
         )
     return None
 
@@ -820,7 +844,7 @@ def selection_from_artifact(
     # A slim selection carries its pairing so the install wiring and marker know
     # which llama runtime provides the ggml libraries.
     runtime = installed_llama_runtime()
-    if runtime is None or runtime[1] != artifact.get("requires_llama_tag"):
+    if runtime is None or not llama_runtime_pairs(runtime[1], artifact.get("requires_llama_tag")):
         raise PrebuiltFallback(
             "the paired llama.cpp runtime changed underneath the slim whisper selection"
         )

tests/studio/install/test_install_whisper_prebuilt_logic.py

@@ -441,8 +441,9 @@ def test_main_forwards_requested_whisper_tags(tmp_path, monkeypatch):
     monkeypatch.setattr(
         M,
         "resolve_prebuilt",
-        lambda host, **kwargs: seen.update(kwargs)
-        or {"prebuilt_available": False, "repo": "unslothai/whisper.cpp"},
+        lambda host, **kwargs: (
+            seen.update(kwargs) or {"prebuilt_available": False, "repo": "unslothai/whisper.cpp"}
+        ),
     )
     assert M.main(["--resolve-prebuilt", "v1.8.0", "--output-format", "json"]) == 0
     assert seen["whisper_tag"] == "v1.8.0"
@@ -807,6 +808,50 @@ def test_slim_release_tag_skew_has_distinct_compatibility_error(tmp_path, monkey
         M.select_artifact_with_fallback(manifest, _cuda_host(), "cuda")
 
 
+# A newer llama build that keeps the same ggml commit as SLIM_LLAMA_TAG.
+NEWER_LLAMA_TAG = "b10079-mix-fb3d4ca"
+
+
+@pytest.mark.parametrize(
+    "installed,required,pairs",
+    [
+        (SLIM_LLAMA_TAG, SLIM_LLAMA_TAG, True),  # exact tag
+        (NEWER_LLAMA_TAG, SLIM_LLAMA_TAG, True),  # newer build, same ggml commit
+        ("b10069-mix-0000000", SLIM_LLAMA_TAG, False),  # same build, different ggml
+        (SLIM_LLAMA_TAG, None, False),  # no requirement recorded
+        ("b10069", "b10069", True),  # tag without -mix-, exact only
+        ("b10070", "b10069", False),  # tag without -mix-, no shared key
+    ],
+)
+def test_llama_runtime_pairs_keys_on_ggml_commit(installed, required, pairs):
+    assert M.llama_runtime_pairs(installed, required) is pairs
+
+
+def test_slim_pairs_across_llama_build_bump_with_same_ggml(tmp_path, monkeypatch):
+    # The live failure: the llama installer advances to a newer build that keeps
+    # the same ggml commit, so the slim bundle's paired runtime is ABI-identical
+    # and must still select rather than degrade to CPU or report unavailable.
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, NEWER_LLAMA_TAG, "cuda13-newer")
+    )
+    artifact, backend, used_fallback = M.select_artifact_with_fallback(
+        _slim_manifest(), _cuda_host(), "cuda"
+    )
+    assert artifact["asset"] == SLIM_ASSET
+    assert backend == "cuda" and used_fallback is False
+
+
+def test_slim_build_bump_same_ggml_is_not_a_compatibility_error(tmp_path, monkeypatch):
+    # A same-ggml build bump must not surface as a release incompatibility (the
+    # update path reports that as unavailable); only a real ggml skew does.
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, NEWER_LLAMA_TAG, "cuda13-newer")
+    )
+    assert M._slim_release_incompatibility(_slim_manifest(), _cuda_host()) is None
+
+
 def test_link_ggml_runtime_hardlinks_every_ggml_library(tmp_path):
     bin_dir = _fake_llama_bin(tmp_path)
     whisper_bin = tmp_path / "whisper.cpp" / "build" / "bin"