feat: add Granite 4.2 model defaults

.github/workflows/quality.yml

@@ -105,29 +105,38 @@ jobs:
       - name: Download NLTK data
         run: uv run python -m nltk.downloader punkt_tab
       - name: Install Ollama
-        # Pinned to a GCC-13 build to fix a segfault loading granite4.1:3b on
-        # AMX-capable ubuntu-latest runners (#1388). Ollama's Linux builds through
-        # 0.32.1 were compiled with GCC 11, which miscompiles the AMX CPU kernels in
-        # libggml-cpu-sapphirerapids.so -- so ~1/3 of runs (the AMX-capable runner
-        # draws) crashed during model warmup, independent of ollama version and
-        # unreachable by any env var (the code path is correct; the machine code was
-        # not). Ollama PR #17244 ("bump Linux toolchain to GCC 13", Fixes ollama#17006
-        # / #17205) first shipped in the 0.32.2 line. The rc0 tag has since been
-        # removed (promoted to the v0.32.2 release), so its download URL now 404s and
-        # the pin is bumped to 0.32.2. Drop this pin once a stable (non-prerelease)
-        # release includes the GCC-13 build. NB: OLLAMA_VERSION must omit the leading
-        # "v" -- install.sh prepends it (a "v" yields a broken .../download/vv0.32.2/...
-        # path).
-        run: curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.32.2 sh
+        # GCC-13 build: fixes a segfault loading Granite models on AMX-capable
+        # runners (Ollama miscompiled AMX kernels through 0.32.1, #1388).
+        # Pinned to 0.33.1 (current stable with the fix); also required for
+        # the reasoning_effort="none" thinking-off support used below.
+        # NB: OLLAMA_VERSION must omit the leading "v" (install.sh prepends it).
+        run: curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.33.1 sh
       - name: Start serving ollama
-        run: nohup ollama serve &
+        # Debug + per-request logging, dumped on failure below: needed to tell
+        # "server hung" from "bad request" when a live-model test stalls.
+        run: OLLAMA_DEBUG=1 OLLAMA_DEBUG_LOG_REQUESTS=1 nohup ollama serve > /tmp/ollama-serve.log 2>&1 &
+      - name: Start resource sampler
+        # Memory/swap/CPU every 30s, dumped on failure below — rules out
+        # memory pressure as a cause of live-model stalls.
+        run: |
+          if command -v vmstat >/dev/null; then
+            (vmstat 30 480) > /tmp/ci-resource-sampler.log 2>&1 &
+          else
+            (for i in $(seq 1 480); do
+               echo "sample $i $(date -u +%FT%TZ)"
+               free -m
+               cat /proc/loadavg
+               echo ---
+               sleep 30
+             done) > /tmp/ci-resource-sampler.log 2>&1 &
+          fi
       - name: Pull models
         # granite-vision-4.1 is not in the Ollama library, so it is pulled from
         # IBM's official GGUF repo on Hugging Face (that build ships the mmproj
         # projector Ollama needs for image input). Drop the hf.co prefix if the
         # model is ever published to the Ollama library directly.
         run: |
-          for model in granite4.1:3b hf.co/ibm-granite/granite-vision-4.1-4b-GGUF:Q4_K_M; do
+          for model in granite4.2:3b granite4:micro-h hf.co/ibm-granite/granite-vision-4.1-4b-GGUF:Q4_K_M; do
             pulled=false
             for i in 1 2 3 4 5; do
               ollama pull "$model" && { pulled=true; break; }
@@ -144,6 +153,20 @@ jobs:
           # The names below are what the tests match against; print them so a
           # tag-name mismatch is diagnosable from the log alone.
           ollama list
+      - name: Constrain the granite4.2:3b context for the CI runner
+        # The published tag ships num_ctx=131072, a ~6 GB KV cache at load —
+        # too much for the 16 GB runner and a cause of the CI stalls. Re-point
+        # the local tag at an 8192-context build of the same weights (same tag
+        # name, so no test code changes; 2x what other Ollama tests use).
+        run: |
+          ollama cp granite4.2:3b granite4.2:3b-128k
+          printf 'FROM granite4.2:3b-128k\nPARAMETER num_ctx 8192\n' \
+            > /tmp/MODELFILE-granite42-ci
+          ollama create granite4.2:3b -f /tmp/MODELFILE-granite42-ci
+          # grep exits non-zero (failing the step) if num_ctx did not land at
+          # exactly 8192 — matching only '^PARAMETER' would still pass on the
+          # base tag's untouched 131072 value if `ollama create` silently no-oped.
+          ollama show granite4.2:3b --modelfile | grep -qx 'PARAMETER num_ctx 8192'
       - name: Run Tests
         id: tests
         env:
@@ -156,6 +179,31 @@ jobs:
       - name: Send failure message tests
         if: failure() # This step will only run if a previous step failed
         run: echo "Tests failed. Please verify that tests are working locally."
+      - name: Dump Ollama server log on failure
+        if: failure() # Only the test step can produce a useful server log here.
+        # Raw log is too verbose to dump whole; extract requests/errors/model
+        # events plus a short raw tail.
+        run: |
+          echo "=== ollama serve log: requests (GIN), errors, model events ==="
+          grep -iE 'GIN|error|warn|panic|loading|unload' /tmp/ollama-serve.log | tail -n 500 || echo "(no matching lines)"
+          echo "=== ollama serve log (last 50 raw lines) ==="
+          tail -n 50 /tmp/ollama-serve.log || echo "(no ollama serve log found)"
+          echo "=== resource sampler (30 s: memory, swap si/so, CPU) ==="
+          cat /tmp/ci-resource-sampler.log || echo "(no sampler log found)"
+          echo "=== ollama request debug logs (request bodies + replay curls) ==="
+          reqdir=$(grep -oE '/tmp/ollama-request-logs-[0-9a-z]+' /tmp/ollama-serve.log | head -1)
+          if [ -n "$reqdir" ] && [ -d "$reqdir" ]; then
+            echo "file count: $(find "$reqdir" -maxdepth 1 -type f | wc -l)"
+            # Newest 20 files (mtime-descending, "%T@" epoch + path).
+            find "$reqdir" -maxdepth 1 -type f -printf '%T@ %p\n' \
+              | sort -rn | head -20 | while IFS= read -r line; do
+                f=${line#* }
+                echo "--- $f ---"
+                cat "$f"
+              done
+          else
+            echo "(no request debug dir found in serve log)"
+          fi
       - name: Write job summary
         if: always()
         run: |

docs/docs/examples/index.md

@@ -130,4 +130,4 @@ uv run docs/examples/<folder>/<file>.py
 
 **Default backend:** `start_session()` with no arguments connects to a local
 [Ollama](https://ollama.ai) instance running **IBM Granite 4 Micro**
-(`granite4.1:3b`). Make sure Ollama is running before you execute any example.
+(`granite4.2:3b`). Make sure Ollama is running before you execute any example.

docs/docs/getting-started/installation.md

@@ -62,5 +62,5 @@ The default session connects to [Ollama](https://ollama.ai) running locally.
 Install Ollama and pull the default model before running any examples:
 
 ```bash
-ollama pull granite4.1:3b
+ollama pull granite4.2:3b
 ```

docs/docs/getting-started/quickstart.md

@@ -10,7 +10,7 @@ description: "Run your first generative program in minutes."
 ## Hello world
 
 By default, `start_session()` connects to Ollama and uses **IBM Granite 4 Micro**
-(`granite4.1:3b`). Make sure Ollama is running before you run this:
+(`granite4.2:3b`). Make sure Ollama is running before you run this:
 
 ```python
 import mellea
@@ -205,7 +205,7 @@ Hugging Face, and WatsonX are also supported. See
 
 ## Troubleshooting
 
-**`granite4.1:3b` not found** — run `ollama pull granite4.1:3b` before starting.
+**`granite4.2:3b` not found** — run `ollama pull granite4.2:3b` before starting.
 
 **Python 3.13 `outlines` install failure** — `outlines` requires a Rust compiler.
 Either install [Rust](https://www.rust-lang.org/tools/install) or pin Python to 3.12.

docs/docs/troubleshooting/common-errors.md

@@ -6,16 +6,16 @@ description: "Common errors, diagnostic steps, and fixes for Mellea programs."
 
 ## Installation
 
-### `granite4.1:3b` not found
+### `granite4.2:3b` not found
 
 ```text
-Error: model "granite4.1:3b" not found
+Error: model "granite4.2:3b" not found
 ```
 
 Pull the model before running:
 
 ```bash
-ollama pull granite4.1:3b
+ollama pull granite4.2:3b
 ```
 
 ### Python 3.13: `outlines` install failure

docs/docs/troubleshooting/faq.md

@@ -38,7 +38,7 @@ m = MelleaSession(
 )
 ```
 
-## How do I use a model other than `granite4.1:3b`?
+## How do I use a model other than `granite4.2:3b`?
 
 Pass the `model_id` parameter to `start_session()`:
 

docs/docs/tutorials/01-your-first-generative-program.md

@@ -20,7 +20,7 @@ By the end you will have covered:
 > see [Tutorial 03: Using Generative Stubs](../tutorials/using-generative-stubs).
 
 **Prerequisites:** [Quick Start](../getting-started/quickstart) complete,
-Mellea installed (`uv add mellea`), Ollama running locally with `granite4.1:3b` downloaded.
+Mellea installed (`uv add mellea`), Ollama running locally with `granite4.2:3b` downloaded.
 
 ---
 

docs/docs/tutorials/02-streaming-and-async.md

@@ -16,7 +16,7 @@ By the end you will have covered:
 - Context behaviour with concurrent async calls
 
 **Prerequisites:** [Tutorial 01](./your-first-generative-program) complete,
-`pip install mellea`, Ollama running locally with `granite4.1:3b` downloaded.
+`pip install mellea`, Ollama running locally with `granite4.2:3b` downloaded.
 
 ---
 

docs/docs/tutorials/03-using-generative-stubs.md

@@ -16,7 +16,7 @@ By the end you will have covered:
 - Precondition and postcondition validation patterns
 
 **Prerequisites:** [Tutorial 01](./your-first-generative-program) complete,
-`pip install mellea`, Ollama running locally with `granite4.1:3b` downloaded.
+`pip install mellea`, Ollama running locally with `granite4.2:3b` downloaded.
 
 ---
 

mellea/backends/litellm.py

@@ -76,7 +76,10 @@ class LiteLLMBackend(FormatterBackend):
             `ollama_chat/` → localhost:11434, `anthropic/` → Anthropic API).
             Use `None` for cloud providers; set explicitly for local servers
             such as vLLM or a non-default Ollama port.
-        model_options (dict | None): Default model options for generation requests.
+        model_options (dict | None): Default model options applied to every
+            generation request. Per-call options take precedence. Use
+            `{ModelOption.THINKING: False}` here to suppress the think block
+            on models that enable it by default.
 
     Attributes:
         to_mellea_model_opts_map (dict): Mapping from backend-specific option names to
@@ -87,7 +90,7 @@ class LiteLLMBackend(FormatterBackend):
 
     def __init__(
         self,
-        model_id: str = "ollama_chat/" + str(model_ids.IBM_GRANITE_4_1_3B.ollama_name),
+        model_id: str = "ollama_chat/" + str(model_ids.IBM_GRANITE_4_2_3B.ollama_name),
         formatter: ChatFormatter | None = None,
         base_url: str | None = None,
         model_options: dict | None = None,
@@ -384,9 +387,9 @@ async def _generate_from_chat_context_standard(
 
         # Map THINKING to the correct backend parameter(s). Two mechanisms:
         # - chat_template_kwargs.enable_thinking: vLLM/Qwen3/Gemma4 (bool toggle)
-        # - reasoning_effort: LiteLLM/OpenAI-compatible (string level, or True → "medium")
-        # Both are set for True so each server picks up whichever it understands.
-        # NOTE: don't pass reasoning_effort=False — it is invalid; absence disables reasoning.
+        # - reasoning_effort: LiteLLM/OpenAI-compatible (string level; True → "medium",
+        #   False → "none")
+        # Both are set so each server picks up whichever it understands.
         thinking = model_opts.get(ModelOption.THINKING, None)
         original_thinking = thinking  # preserve raw caller value for the generate log
         reasoning_params: dict[str, Any] = {}
@@ -399,8 +402,15 @@ async def _generate_from_chat_context_standard(
                 extra_params["extra_body"] = ctk_body
                 if thinking:
                     reasoning_params["reasoning_effort"] = "medium"
-                # False: do not send reasoning_effort — absent param disables reasoning;
-                # passing False would be invalid.
+                elif "ollama" in self._model_id.split("/")[0]:
+                    # Ollama-served thinking models (e.g. granite4.2) default to
+                    # thinking ON when reasoning_effort is absent; "none" is the
+                    # OpenAI enum value their /v1 endpoint maps to think=false
+                    # (Ollama >= 0.33.1). Real OpenAI/other reasoning providers
+                    # reject "none", so this is scoped to Ollama-routed models
+                    # (same provider-prefix check used for the streaming
+                    # tool-call workaround above).
+                    reasoning_params["reasoning_effort"] = "none"
             else:
                 reasoning_params["reasoning_effort"] = thinking
 

mellea/backends/model_ids.py

@@ -115,6 +115,25 @@ class ModelIdentifier:
     context_length=131072,
 )
 
+# Granite 4.2 Dense Models
+IBM_GRANITE_4_2_3B = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.2-3b",
+    ollama_name="granite4.2:3b",
+    context_length=131072,
+)
+
+IBM_GRANITE_4_2_8B = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.2-8b",
+    ollama_name="granite4.2:8b",
+    context_length=131072,
+)
+
+IBM_GRANITE_4_2_30B = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.2-30b",
+    ollama_name="granite4.2:30b",
+    context_length=131072,
+)
+
 IBM_GRANITE_GUARDIAN_4_1_8B = ModelIdentifier(
     hf_model_name="ibm-granite/granite-guardian-4.1-8b", context_length=131072
 )

mellea/backends/model_options.py

@@ -65,8 +65,11 @@ class ModelOption:
       silently ignored by those servers.
     * `False` — native Ollama backend: sends `think=False`.
       OpenAI-compatible backends: sets `chat_template_kwargs.enable_thinking=False`
-      to suppress the think block. `reasoning_effort` is not sent (passing
-      `False` would be an invalid value for OpenAI; absence disables reasoning).
+      and `reasoning_effort="none"` to suppress the think block. Both are sent so
+      each server type picks up the mechanism it understands: vLLM honours
+      `chat_template_kwargs`, while Ollama's /v1 endpoint (>= 0.33.1) honours
+      `reasoning_effort` — and for models that default to thinking on (e.g.
+      granite4.2) absent `reasoning_effort` means thinking stays on.
     * `"low"` / `"medium"` / `"high"` — passed directly as `reasoning_effort`
       (OpenAI-compatible backends only; no-op on vLLM).
 

mellea/backends/ollama.py

@@ -174,7 +174,7 @@ class OllamaModelBackend(FormatterBackend):
 
     def __init__(
         self,
-        model_id: str | ModelIdentifier = model_ids.IBM_GRANITE_4_1_3B,
+        model_id: str | ModelIdentifier = model_ids.IBM_GRANITE_4_2_3B,
         formatter: ChatFormatter | None = None,
         base_url: str | None = None,
         model_options: dict | None = None,

mellea/backends/openai.py

@@ -519,6 +519,47 @@ def _make_backend_specific_and_remove(
 
         return model_opts
 
+    def _map_thinking_option(
+        self, thinking: Any, extra_body: dict[str, Any]
+    ) -> dict[str, Any]:
+        """Maps `ModelOption.THINKING` to the correct backend parameter(s).
+
+        Two mechanisms, both set (when applicable) so the right server picks
+        up whichever it understands:
+          - `extra_body["chat_template_kwargs"]["enable_thinking"]`: vLLM/Qwen3
+          - `reasoning_effort`: OpenAI/DeepSeek/Ollama (string level; True →
+            "medium", False → "none")
+
+        Ollama-served models (e.g. granite4.2) think by default unless
+        `reasoning_effort="none"` is sent (Ollama >= 0.33.1); real OpenAI
+        rejects `"none"`, so that value is scoped to non-OpenAI servers.
+
+        Args:
+            thinking: the raw `ModelOption.THINKING` value (bool, string
+                reasoning-effort level, or None).
+            extra_body: the in-progress `extra_body` dict for this request;
+                mutated in place to add `chat_template_kwargs` if `thinking`
+                is a bool.
+
+        Returns:
+            dict[str, Any]: `reasoning_effort` params to merge into the
+            request's top-level kwargs, or `{}` if `thinking` is None.
+        """
+        reasoning_params: dict[str, Any] = {}
+        if thinking is None:  # False is a valid value — cannot use `if thinking`
+            return reasoning_params
+        if type(thinking) is bool:
+            ctk = extra_body.get("chat_template_kwargs", {}) or {}
+            ctk["enable_thinking"] = thinking
+            extra_body["chat_template_kwargs"] = ctk
+            if thinking:
+                reasoning_params["reasoning_effort"] = "medium"
+            elif self._server_type != _ServerType.OPENAI:
+                reasoning_params["reasoning_effort"] = "none"
+        else:
+            reasoning_params["reasoning_effort"] = thinking
+        return reasoning_params
+
     def _merge_user_extra_body(
         self, base: dict[str, Any], user: dict[str, Any] | None
     ) -> dict[str, Any]:
@@ -809,22 +850,8 @@ async def _generate_from_intrinsic(
         user_extra_body = user_api_params.pop("extra_body", None)
         api_params.update(user_api_params)
 
-        # Map THINKING to the correct backend parameter(s). Two mechanisms:
-        # - chat_template_kwargs.enable_thinking: vLLM/Qwen3 (bool toggle)
-        # - reasoning_effort: OpenAI/DeepSeek (string level, or True → "medium")
-        # Both are set for True so the right server picks up whichever it understands.
         thinking = model_options.get(ModelOption.THINKING)
-        if thinking is not None:  # False is a valid value — cannot use `if thinking`
-            if type(thinking) is bool:
-                ctk = extra_body.get("chat_template_kwargs", {}) or {}
-                ctk["enable_thinking"] = thinking
-                extra_body["chat_template_kwargs"] = ctk
-                if thinking:
-                    api_params["reasoning_effort"] = "medium"
-                # False: don't send reasoning_effort — OpenAI disables reasoning by
-                # default when the param is absent; passing False would be invalid.
-            else:
-                api_params["reasoning_effort"] = thinking
+        api_params.update(self._map_thinking_option(thinking, extra_body))
 
         extra_body = self._merge_user_extra_body(extra_body, user_extra_body)
 
@@ -1048,25 +1075,11 @@ async def _generate_from_chat_context_standard(
         formatted_tools = convert_tools_to_json(tools)
         use_tools = len(formatted_tools) > 0
 
-        # Map THINKING to the correct backend parameter(s). Two mechanisms:
-        # - chat_template_kwargs.enable_thinking: vLLM/Qwen3 (bool toggle)
-        # - reasoning_effort: OpenAI/DeepSeek (string level, or True → "medium")
-        # NOTE: don't pass reasoning_effort to non-reasoning models (e.g. gpt-4o).
+        # NOTE: don't pass THINKING to non-reasoning models (e.g. gpt-4o).
         thinking = model_opts.get(ModelOption.THINKING)
-        reasoning_params: dict[str, Any] = {}
-        if thinking is not None:  # False is a valid value — cannot use `if thinking`
-            if type(thinking) is bool:
-                ctk_body: dict[str, Any] = extra_params.get("extra_body", {}) or {}
-                ctk = ctk_body.get("chat_template_kwargs", {}) or {}
-                ctk["enable_thinking"] = thinking
-                ctk_body["chat_template_kwargs"] = ctk
-                extra_params["extra_body"] = ctk_body
-                if thinking:
-                    reasoning_params["reasoning_effort"] = "medium"
-                # False: don't send reasoning_effort — OpenAI disables reasoning by
-                # default when the param is absent; passing False would be invalid.
-            else:
-                reasoning_params["reasoning_effort"] = thinking
+        ctk_body: dict[str, Any] = extra_params.get("extra_body", {}) or {}
+        reasoning_params = self._map_thinking_option(thinking, ctk_body)
+        extra_params["extra_body"] = ctk_body
 
         # Request usage information in streaming responses
         if model_opts.get(ModelOption.STREAM, False):