mtplx 2.7.0__tar.gz → 2.7.2__tar.gz

{mtplx-2.7.0 → mtplx-2.7.2}/CHANGELOG.md

@@ -4,6 +4,83 @@ All notable user-facing changes to MTPLX. The format is based on
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow
 [Semantic Versioning](https://semver.org/).
 
+## [2.7.2] - 2026-08-16
+
+An emergency fix for `mtplx pull`: on 2.7.1 and older, re-pulling a repo that
+changed upstream — such as the Qwen 3.8 repos re-published on 2026-08-15 with
+their vision towers restored — can corrupt the local copy. Upgrade before
+pulling.
+
+### Fixed
+
+- **The Qwen 3.8 Hugging Face artifacts can see again (#263).** All six
+  published 3.8 repos (Bare Speed, Optimized Speed, Optimized Quality and
+  their FP16 siblings) shipped without their vision towers: the forge
+  `mlx_lm.convert` lane serializes only the text model, so the 333
+  `model.visual.*` tensors, `vision_config`, and the preprocessor sidecars
+  were silently dropped. The repos were re-published on 2026-08-15 with the
+  official bf16 tower grafted back as an index-registered
+  `model-vision.safetensors` (byte-for-byte from `Qwen/Qwen3.8-27B`; language
+  and MTP tensors untouched). Existing installs pick the delta up with
+  `mtplx pull`; images, `/health` `vision.enabled`, prompt caching with
+  images, and MTP decode were verified on all six builds. Thanks to
+  @kjellix for the report and the proven graft procedure.
+- **Forge keeps vision towers from now on.** `mtplx forge build` grafts the
+  source's vision tower, `vision_config`, and preprocessor sidecars into the
+  converted artifact on every lane (`mtplx/vision_graft.py`), and fails
+  closed when a source that declares `vision_config` would produce a blind
+  artifact. A repair script for already-forged artifacts ships as
+  `scripts/graft_vision_tower.py`, and the pillar gate now asserts
+  `vision.enabled` before the vision-cache check so a blind build fails
+  loudly with the real cause.
+- **`mtplx pull` no longer corrupts files that changed upstream (#258,
+  #234).** The
+  progress downloader treated a size-mismatched *complete* local file as a
+  resumable partial and byte-range-appended the remote tail onto the old
+  content — updating a repo in place (for example the restored 3.8 vision
+  indexes) corrupted `config.json` and `model.safetensors.index.json` and
+  left the local copy unloadable until re-downloaded. Stale files are now
+  discarded and re-fetched whole; genuine `*.incomplete` partials still
+  resume. **Users on ≤2.7.1 should upgrade before pulling repaired repos**;
+  a failed pull from an older build is recovered by deleting the corrupt
+  `config.json` + `model.safetensors.index.json` and pulling again.
+
+## [2.7.1] - 2026-08-15
+
+Clears the 2.7.0 known-issues list: `xhigh` is now selectable everywhere it is
+offered, and the app's KV cache quantization toggle reaches Qwen 3.8.
+
+### Fixed
+
+- **`xhigh` no longer bounces back to `medium` in the app.** Choosing it in
+  Inference settings while the model was running, or running
+  `mtplx config set reasoning_effort xhigh`, was rejected: those two writing
+  surfaces carried their own hardcoded effort list that stopped at `high`,
+  while the request path already knew `xhigh`. Because the live-settings POST
+  is all-or-nothing, one unknown level threw away the whole payload and the
+  picker snapped back to the family default. All four writing surfaces
+  (serve/CLI `--reasoning-effort`, `mtplx config set`, the live-settings POST)
+  now read one shared vocabulary in `mtplx/reasoning_effort.py`, and narrowing
+  to what the loaded model actually supports stays where it belongs — the
+  family's own `effort_levels`.
+- **The app's KV cache quantization toggle now applies to Qwen 3.8.** The
+  launch gate listed only `qwen3_5`/`qwen3_6`, so a 3.8 launch exported
+  nothing while Settings showed `q8`.
+- **`mtplx doctor` names the model it actually checks** instead of hardcoding
+  the previous default, and turbo's profile note reports the real compiled
+  verify fence (32,768, raised from 12,288 in 2.7.0) rather than the stale one.
+- **A built app can no longer rank below the release it supersedes.** Sparkle
+  orders updates by `CFBundleVersion` alone, and the number derived from the
+  version was narrow enough that 2.7.1 computed lower than the 2.7.0 that
+  shipped — so a fresh build offered 2.7.0 to itself as an update. The
+  derivation now has room for the whole version and the appcast reads its
+  number back off the built bundle instead of being told one separately.
+
+### Still open from 2.7.0
+
+- With reasoning switched off in a plain chat with no tools, Qwen 3.8 can
+  still emit a stray tool call and cut the turn short. Keep thinking on.
+
 ## [2.7.0] - 2026-08-15
 
 Qwen3.8 support 🎉. Qwen3.8-27B came out on 2026-08-14; this release runs it
@@ -1338,6 +1415,8 @@ working as one product. Full notes:
   completions, and Anthropic `stop_sequences`) and `/v1/completions`
   streams tokens as they are generated with real finish reasons.
 
+[2.7.2]: https://github.com/youssofal/MTPLX/releases/tag/v2.7.2
+[2.7.1]: https://github.com/youssofal/MTPLX/releases/tag/v2.7.1
 [2.7.0]: https://github.com/youssofal/MTPLX/releases/tag/v2.7.0
 [2.6.0]: https://github.com/youssofal/MTPLX/releases/tag/v2.6.0
 [2.5.4]: https://github.com/youssofal/MTPLX/releases/tag/v2.5.4

{mtplx-2.7.0 → mtplx-2.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mtplx
-Version: 2.7.0
+Version: 2.7.2
 Summary: Native MTP speculative decoding for Qwen3-Next on Apple Silicon.
 Author-email: Youssof Altoukhi <business@youssofal.com>
 License-Expression: Apache-2.0

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/backends/descriptors.py

@@ -82,6 +82,10 @@ class ReasoningCodec:
     supported: bool = True
     modes: tuple[str, ...] = ("auto", "on", "off")
     history_policy: str = "preserve_when_enabled"
+    # Drawn from mtplx.reasoning_effort.REASONING_EFFORT_LEVELS: the app
+    # renders these verbatim, so a level outside that vocabulary is one the
+    # user can pick and no writing surface will accept
+    # (test_reasoning_effort_vocabulary_covers_every_family pins it).
     effort_levels: tuple[str, ...] = ()
     default_effort: str | None = None
 

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/cli.py

@@ -23,6 +23,7 @@ from .profiles import (
     list_profiles,
     resolve_profile_name,
 )
+from .reasoning_effort import REASONING_EFFORT_CHOICES
 from .runtime_options import canonicalize_flag_tokens, normalize_paged_kv_quantization
 from .version import DISPLAY_VERSION, __version__
 
@@ -573,7 +574,7 @@ def _add_reasoning_arg(
 def _add_reasoning_effort_arg(parser: argparse.ArgumentParser) -> None:
     parser.add_argument(
         "--reasoning-effort",
-        choices=["auto", "low", "medium", "high", "xhigh"],
+        choices=list(REASONING_EFFORT_CHOICES),
         default="auto",
         help=(
             "Reasoning effort for models that expose levels, such as Qwen 3.8 "

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/commands/forge.py

@@ -1028,6 +1028,9 @@ def _cmd_build(args: Any) -> int:
     if _cancel_requested(args.run_id):
         raise ForgeError("forge cancelled", code=130)
 
+    _ensure_vision_tower(source_path, destination)
+    _validate_vision_payload(source_path, destination)
+
     _calibrate_sidecar(
         source_path,
         destination,
@@ -1767,6 +1770,51 @@ def _validate_mtp_sidecar_payload(destination: Path) -> dict[str, Any] | None:
     return audit
 
 
+def _ensure_vision_tower(source: Path, destination: Path) -> None:
+    """Restore the vision tower that text-only convert lanes drop.
+
+    ``mlx_lm.convert`` serializes only the text model, so multimodal sources
+    lose their vision tensors, ``vision_config``, and preprocessor sidecars
+    (issue #263). Graft them back from the source directory. A no-op for
+    text-only sources and for lanes (mirror, compressed-tensors) that already
+    preserved the tower.
+    """
+    if not source.is_dir():
+        return
+    from mtplx.vision_graft import VisionGraftError, graft_vision_tower
+
+    try:
+        report = graft_vision_tower(source, destination, verify_load=True)
+    except VisionGraftError as exc:
+        raise ForgeError(f"vision tower graft failed: {exc}") from exc
+    if report.get("status") == "grafted":
+        _err(
+            f"[forge] restored vision tower: {report.get('tensors')} tensors "
+            f"({report.get('bytes')} bytes) in {report.get('vision_file')}"
+        )
+
+
+def _validate_vision_payload(source: Path, destination: Path) -> None:
+    """Fail closed when a multimodal source produced a blind artifact."""
+    if not source.is_dir():
+        return
+    try:
+        source_config = _load_json(source / "config.json")
+    except Exception:
+        return
+    if not isinstance(source_config, dict) or not isinstance(
+        source_config.get("vision_config"), dict
+    ):
+        return
+    from mtplx.vision import vision_spec_for_model_dir
+
+    if vision_spec_for_model_dir(destination) is None:
+        raise ForgeError(
+            "source declares vision_config but the forged artifact resolves no "
+            "vision tower; the convert lane dropped it (issue #263)"
+        )
+
+
 def _audit_mtp_sidecar_payload(mtp_path: Path) -> dict[str, Any]:
     problems: list[str] = []
     try:
@@ -2922,6 +2970,9 @@ def _stamp_runtime_metadata(
         metadata["mtp_sidecar"] = inspection.mtp.sidecar_format
     else:
         metadata.setdefault("mtp_sidecar", "mtp.safetensors")
+    vision_stamp = _vision_metadata_stamp(model_path)
+    if vision_stamp is not None:
+        metadata["vision"] = vision_stamp
     metadata["base_trunk"] = source_repo
     metadata.setdefault("artifact_role", "forge-local")
     metadata["forge_provenance"] = {
@@ -2939,6 +2990,32 @@ def _stamp_runtime_metadata(
     return metadata
 
 
+def _vision_metadata_stamp(model_path: Path) -> dict[str, Any] | None:
+    """Provenance for artifacts that carry a vision tower, or None."""
+    from mtplx.vision import resolve_vision_prefix
+
+    index_path = model_path / "model.safetensors.index.json"
+    if not index_path.exists():
+        return None
+    try:
+        index = _load_json(index_path)
+    except Exception:
+        return None
+    weight_map = index.get("weight_map") if isinstance(index, dict) else None
+    if not isinstance(weight_map, dict):
+        return None
+    prefix = resolve_vision_prefix(weight_map)
+    if prefix is None:
+        return None
+    vision_keys = [key for key in weight_map if str(key).startswith(prefix)]
+    shards = sorted({str(weight_map[key]) for key in vision_keys})
+    return {
+        "tensor_count": len(vision_keys),
+        "prefix": prefix,
+        "shards": shards,
+    }
+
+
 def _speed_evidence(rows: list[dict[str, Any]]) -> dict[str, Any]:
     winner = _winning_row(rows)
     ar_row = next((row for row in rows if row.get("depth") == 0), None)

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/commands/public.py

@@ -125,6 +125,7 @@ from mtplx.profiles import (
     restore_profile_env,
     runtime_env_with_contract_overrides,
 )
+from mtplx.reasoning_effort import REASONING_EFFORT_CHOICES
 from mtplx.server_urls import (
     bind_label,
     connect_host_for_bind,
@@ -14347,8 +14348,10 @@ def cmd_config_public(args: Any) -> int:
         )
     if key == "reasoning" and value not in {"auto", "on", "off"}:
         raise SystemExit("reasoning must be auto, on, or off")
-    if key == "reasoning_effort" and value not in {"auto", "low", "medium", "high"}:
-        raise SystemExit("reasoning_effort must be auto, low, medium, or high")
+    if key == "reasoning_effort" and value not in REASONING_EFFORT_CHOICES:
+        raise SystemExit(
+            "reasoning_effort must be " + ", ".join(REASONING_EFFORT_CHOICES)
+        )
     if key in {
         "max_active_requests",
         "decode_batch_max",

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/diagnostics.py

@@ -26,7 +26,7 @@ from mtplx.profiles import DEFAULT_HF_MODEL_ID, DEFAULT_PROFILE_NAME
 
 
 GIB = 1024**3
-DEFAULT_SPEED_MODEL_SIZE_BYTES = 20_392_433_868  # Qwen3.8-27B Optimized Speed (Hub bytes)
+DEFAULT_SPEED_MODEL_SIZE_BYTES = 21_313_949_792  # Qwen3.8-27B Optimized Speed (Hub bytes, incl. restored vision tower #263)
 MIN_RECOMMENDED_MEMORY_BYTES = 48 * GIB
 SUPPORT_MACOS_MAJOR = 14
 SUPPORT_PYTHON = (3, 11)
@@ -441,7 +441,7 @@ def build_diagnostic_checks(
             },
             "default model available in the HF cache or as the verified local startup model",
             model_cache_fix,
-            "https://huggingface.co/Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed",
+            f"https://huggingface.co/{DEFAULT_HF_MODEL_ID}",
             f"mtplx pull {DEFAULT_HF_MODEL_ID}",
         )
     )
@@ -452,9 +452,9 @@ def build_diagnostic_checks(
             "fail" if stale else "pass",
             "error",
             DEFAULT_HF_MODEL_ID,
-            "Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed",
+            "a published Youssofal/... repo (not a local mtplx/ or models/ path)",
             "Pull the default model, or pass --model to serve a different one.",
-            "https://huggingface.co/Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed",
+            f"https://huggingface.co/{DEFAULT_HF_MODEL_ID}",
             f"mtplx pull {DEFAULT_HF_MODEL_ID}",
         )
     )
@@ -481,7 +481,7 @@ def build_diagnostic_checks(
                 hf_observed,
                 "default Hugging Face repo is reachable from this machine",
                 "Check network/HF auth or verify the model repo is public.",
-                "https://huggingface.co/Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed",
+                f"https://huggingface.co/{DEFAULT_HF_MODEL_ID}",
                 f"mtplx pull {DEFAULT_HF_MODEL_ID}",
             )
         )

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/hf_loader.py

@@ -637,10 +637,12 @@ def _download_repo_file(
 
     partial = target.with_name(target.name + ".incomplete")
     if target.exists():
-        if not partial.exists():
-            target.replace(partial)
-        else:
-            target.unlink()
+        # A size-mismatched final file is a stale version of a file that
+        # changed upstream (e.g. a repaired index gaining vision entries),
+        # not an interrupted download. Resuming from it would append the
+        # remote tail onto old content and corrupt the file, so discard it.
+        # Only a leftover *.incomplete partial may be range-resumed.
+        target.unlink()
     existing = partial.stat().st_size if partial.exists() else 0
     if expected_size is not None and existing > expected_size:
         partial.unlink()

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/model_catalog.py

@@ -128,8 +128,9 @@ OFFICIAL_CATALOG: tuple[CatalogModel, ...] = (
         ),
         hf_model_id="Youssofal/Qwen3.8-27B-MTPLX-Bare-Speed",
         # Exact byte sum of the published HF repo files (2026-08-15 tree API;
-        # three trunk shards + bf16 MTP sidecar + tokenizer + card).
-        size_bytes=16_002_648_138,
+        # three trunk shards + bf16 MTP sidecar + restored bf16 vision tower
+        # (#263) + tokenizer + card).
+        size_bytes=16_924_164_062,
         # Measured 2026-08-14: request-log MLX high-water 19.6 GiB during
         # quiet-window 2.4k-context serving (boot + Flappy arms + rung).
         peak_memory_gib=20.0,
@@ -150,8 +151,9 @@ OFFICIAL_CATALOG: tuple[CatalogModel, ...] = (
         ),
         hf_model_id="Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed",
         # Exact byte sum of the published HF repo files (2026-08-15 tree API;
-        # module_overrides recipe, 5.807 bits/weight, bf16 MTP sidecar).
-        size_bytes=20_392_433_868,
+        # module_overrides recipe, 5.807 bits/weight, bf16 MTP sidecar,
+        # restored bf16 vision tower (#263)).
+        size_bytes=21_313_949_792,
         # Measured 2026-08-14: request-log MLX high-water 24.6 GiB during
         # quiet-window 2.4k-context serving (boot + Flappy arms + rung).
         peak_memory_gib=25.0,
@@ -169,8 +171,9 @@ OFFICIAL_CATALOG: tuple[CatalogModel, ...] = (
             "8-bit dynamic quant. Good coding speeds and perfect quality."
         ),
         hf_model_id="Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality",
-        # Exact byte sum of the published HF repo files (2026-08-15 tree API).
-        size_bytes=29_449_324_149,
+        # Exact byte sum of the published HF repo files (2026-08-15 tree API;
+        # includes the restored bf16 vision tower, #263).
+        size_bytes=30_370_840_073,
         # Measured 2026-08-14: request-log MLX high-water 32.9 GiB during
         # quiet-window 2.4k-context serving (boot + Flappy arms + rung).
         peak_memory_gib=33.0,
@@ -193,8 +196,9 @@ OFFICIAL_CATALOG: tuple[CatalogModel, ...] = (
             "coding tasks. FP16 build for M1 and M2 Macs."
         ),
         hf_model_id="Youssofal/Qwen3.8-27B-MTPLX-Bare-Speed-FP16",
-        # Exact byte sum of the local sibling at build time (2026-08-15).
-        size_bytes=16_003_127_584,
+        # Exact byte sum of the published HF repo files (2026-08-15 tree API;
+        # includes the restored bf16 vision tower, #263).
+        size_bytes=16_924_647_669,
         # Same packs and tensor bytes as the parent; peak carried over.
         peak_memory_gib=20.0,
         recommended_tiers=frozenset({LEGACY_TIER}),
@@ -213,8 +217,9 @@ OFFICIAL_CATALOG: tuple[CatalogModel, ...] = (
             "FP16 build for M1 and M2 Macs. Recommended."
         ),
         hf_model_id="Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed-FP16",
-        # Exact byte sum of the local sibling at build time (2026-08-15).
-        size_bytes=20_392_914_234,
+        # Exact byte sum of the published HF repo files (2026-08-15 tree API;
+        # includes the restored bf16 vision tower, #263).
+        size_bytes=21_314_434_309,
         # Same packs and tensor bytes as the parent; peak carried over.
         peak_memory_gib=25.0,
         recommended_tiers=frozenset({LEGACY_TIER}),
@@ -232,8 +237,9 @@ OFFICIAL_CATALOG: tuple[CatalogModel, ...] = (
             "FP16 build for M1 and M2 Macs."
         ),
         hf_model_id="Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality-FP16",
-        # Exact byte sum of the local sibling at build time (2026-08-15).
-        size_bytes=29_449_805_992,
+        # Exact byte sum of the published HF repo files (2026-08-15 tree API;
+        # includes the restored bf16 vision tower, #263).
+        size_bytes=30_371_326_040,
         # Same packs and tensor bytes as the parent; peak carried over.
         peak_memory_gib=33.0,
         recommended_tiers=frozenset({LEGACY_TIER}),

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/profiles.py

@@ -553,8 +553,9 @@ TURBO_PROFILE = RuntimeProfile(
         "4-bit and 8-bit affine models; 6-bit models silently run the "
         "stock path.",
         "Compiled verify engages on 4-bit and 8-bit affine trunks at "
-        "contexts <= 12288 (parity2-validated on both); other "
-        "quantizations/contexts run the eager verify path unchanged.",
+        "contexts <= 32768 (parity2-validated on both; fence raised from "
+        "12288 in 2.7.0); other quantizations/contexts run the eager verify "
+        "path unchanged.",
         "Measured 2026-07-02/03 on M5 Max chat lane (app-launch flags, "
         "thinking on): 27B Optimized-Speed 44.7 -> 58-60 tok/s (vk_k "
         "within ~2% of the retired dflash-port kernel both directions); "

mtplx-2.7.2/mtplx/reasoning_effort.py

@@ -0,0 +1,37 @@
+"""Canonical MTPLX reasoning-effort vocabulary shared by CLI and server code.
+
+Each model family declares the subset it supports through its
+``ReasoningCodec.effort_levels``, and that subset is exactly what the app's
+effort picker renders. Every *writing* surface — serve/CLI argparse,
+``mtplx config set``, and the live-settings POST — must therefore accept this
+whole vocabulary and leave the per-family narrowing to the one place that
+knows which model is loaded (``_reasoning_effort_for_state``).
+
+Surfaces that inline their own list go stale the moment a family ships a new
+level. That is how 2.7.0 shipped with Qwen 3.8's ``xhigh`` bouncing back to
+``medium``: the request path knew the level, the live-settings coercer did
+not, and because that POST is all-or-nothing the app's whole settings payload
+was rejected.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+# Ordered low to high.
+REASONING_EFFORT_LEVELS = ("low", "medium", "high", "xhigh")
+# "auto" is not a level; it means "use the loaded family's default_effort".
+REASONING_EFFORT_AUTO = "auto"
+REASONING_EFFORT_CHOICES = (REASONING_EFFORT_AUTO, *REASONING_EFFORT_LEVELS)
+
+
+def normalize_reasoning_effort(
+    value: Any, *, default: str = REASONING_EFFORT_AUTO
+) -> str:
+    effort = str(value or default).strip().lower()
+    if effort not in REASONING_EFFORT_CHOICES:
+        raise ValueError(
+            "reasoning_effort must be one of: " + ", ".join(REASONING_EFFORT_CHOICES)
+        )
+    return effort

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/server/openai.py

@@ -83,6 +83,7 @@ from mtplx.mtp_batch_numerics import (
 )
 from mtplx.backends.descriptors import (
     BackendDescriptor,
+    ReasoningCodec,
     assistant_target_distribution_choices,
     descriptor_for_backend_id,
     descriptor_for_model,
@@ -111,6 +112,10 @@ from mtplx.gemma4_pair import (
     resolve_gemma4_pair_paths,
 )
 from mtplx.model_scheduler import ModelWorkScheduler
+from mtplx.reasoning_effort import (
+    REASONING_EFFORT_CHOICES,
+    normalize_reasoning_effort as _normalize_reasoning_effort,
+)
 from mtplx.retrieval import RetrievalError, RetrievalTrustError
 from mtplx.sampling import SamplerConfig
 from mtplx.profiles import (
@@ -13286,12 +13291,7 @@ def _coerce_setting(name: str, value: Any) -> Any:
             )
         return text
     if name == "reasoning_effort":
-        text = str(value).strip().lower()
-        if text not in {"auto", "low", "medium", "high"}:
-            raise ValueError(
-                "reasoning_effort must be 'auto', 'low', 'medium', or 'high'"
-            )
-        return text
+        return _normalize_reasoning_effort(value)
     return value
 
 
@@ -21727,15 +21727,6 @@ def _thinking_enabled_for_request(
     )
 
 
-def _normalize_reasoning_effort(value: Any, *, default: str = "auto") -> str:
-    effort = str(value or default).strip().lower()
-    if effort not in {"auto", "low", "medium", "high", "xhigh"}:
-        raise ValueError(
-            "reasoning_effort must be one of: auto, low, medium, high, xhigh"
-        )
-    return effort
-
-
 def _reasoning_effort_for_state(
     state: ServerState,
     *,
@@ -29022,7 +29013,7 @@ def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
     )
     parser.add_argument(
         "--reasoning-effort",
-        choices=["auto", "low", "medium", "high", "xhigh"],
+        choices=list(REASONING_EFFORT_CHOICES),
         default="auto",
         help=(
             "Backend reasoning effort. Qwen 3.8 exposes xhigh/medium/low "

{mtplx-2.7.0 → mtplx-2.7.2}/mtplx/version.py

@@ -2,5 +2,5 @@
 
 from __future__ import annotations
 
-__version__ = "2.7.0"
-DISPLAY_VERSION = "2.7.0"
+__version__ = "2.7.2"
+DISPLAY_VERSION = "2.7.2"