minder-cli 0.5.3__tar.gz → 0.5.4__tar.gz

{minder_cli-0.5.3 → minder_cli-0.5.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: minder-cli
-Version: 0.5.3
+Version: 0.5.4
 Summary: Minder CLI is the command-line interface for the Minder self-hosted MCP platform.
 Project-URL: Homepage, https://github.com/hiimtrung/minder
 Project-URL: Repository, https://github.com/hiimtrung/minder

{minder_cli-0.5.3 → minder_cli-0.5.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "minder-cli"
-version = "0.5.3"
+version = "0.5.4"
 description = "Minder CLI is the command-line interface for the Minder self-hosted MCP platform."
 readme = "README.md"
 requires-python = ">=3.14"

{minder_cli-0.5.3 → minder_cli-0.5.4}/src/minder/embedding/local.py

@@ -54,9 +54,7 @@ class LocalEmbeddingProvider:
             return
 
         if not llama_cpp_usable():
-            logger.warning(
-                "CPU does not support AVX2; llama.cpp unavailable. Using mock embedding."
-            )
+            logger.warning("llama.cpp not usable on this host; embedding running in mock mode.")
             return
 
         try:

{minder_cli-0.5.3 → minder_cli-0.5.4}/src/minder/llm/llama_cpp_llm.py

@@ -52,9 +52,7 @@ class LlamaCppLLM:
             return
 
         if not llama_cpp_usable():
-            logger.warning(
-                "CPU does not support AVX2; llama.cpp unavailable. Falling back to mock mode."
-            )
+            logger.warning("llama.cpp not usable on this host; LLM running in mock mode.")
             return
 
         cache_key = f"{self._model_repo}:{self._model_file}"

{minder_cli-0.5.3 → minder_cli-0.5.4}/src/minder/runtime.py

@@ -26,18 +26,22 @@ def load_attr(module_name: str, attr_name: str) -> Any | None:
 def llama_cpp_usable() -> bool:
     """Return True only if llama.cpp can actually run on this CPU.
 
-    Some CI runners report AVX2 in /proc/cpuinfo but the hypervisor blocks
-    the instructions, causing an unrecoverable SIGILL.  We probe by running
-    a subprocess that initialises the llama.cpp backend; if the subprocess
-    is killed by SIGILL (returncode == -signal.SIGILL) we mark the library
-    as unavailable and fall back to mock mode.  Result is cached for the
-    lifetime of the process.
+    Some CI runners / VM hypervisors block AVX2 instructions even when
+    /proc/cpuinfo advertises them, causing an unrecoverable SIGILL.  We
+    probe by running a subprocess that initialises the ggml backend; if the
+    subprocess is killed by SIGILL (returncode == -signal.SIGILL) we mark
+    the library as unavailable and fall back to mock mode.
+
+    Build the Docker image with CMAKE_ARGS containing -DGGML_AVX2=OFF (and
+    related flags) to compile a portable binary that avoids SIGILL entirely.
+    Result is cached for the lifetime of the process.
     """
     global _LLAMA_CPP_PROBE
     if _LLAMA_CPP_PROBE is not None:
         return _LLAMA_CPP_PROBE
 
     if not module_available("llama_cpp"):
+        logger.warning("llama-cpp-python is not installed. Falling back to mock mode.")
         _LLAMA_CPP_PROBE = False
         return False
 
@@ -48,19 +52,26 @@ def llama_cpp_usable() -> bool:
                 "-c",
                 # Instantiating Llama triggers ggml_backend_reg_count → ggml_cpu_init.
                 # /dev/null causes a model-load error (returncode 1) on healthy CPUs;
-                # SIGILL kills the process (returncode -4) on unsupported CPUs.
+                # SIGILL (returncode -4) means the CPU doesn't support the compiled ISA.
                 "from llama_cpp import Llama; Llama(model_path='/dev/null', verbose=False)",
             ],
             capture_output=True,
             timeout=30,
         )
-        _LLAMA_CPP_PROBE = proc.returncode != -signal.SIGILL
+        if proc.returncode == -signal.SIGILL:
+            logger.warning(
+                "llama.cpp unavailable: CPU or hypervisor blocked an instruction "
+                "(SIGILL). Rebuild the image with CMAKE_ARGS containing "
+                "-DGGML_AVX2=OFF to fix this. Falling back to mock mode."
+            )
+            _LLAMA_CPP_PROBE = False
+        else:
+            _LLAMA_CPP_PROBE = True
+    except subprocess.TimeoutExpired:
+        logger.warning("llama.cpp usability probe timed out. Falling back to mock mode.")
+        _LLAMA_CPP_PROBE = False
     except Exception as exc:
-        logger.warning("llama.cpp usability probe error: %s", exc)
+        logger.warning("llama.cpp usability probe error: %s. Falling back to mock mode.", exc)
         _LLAMA_CPP_PROBE = False
 
-    if not _LLAMA_CPP_PROBE:
-        logger.warning(
-            "llama.cpp unavailable on this CPU (SIGILL probe). Falling back to mock mode."
-        )
     return _LLAMA_CPP_PROBE