Studio: fix tier detection for models loaded via custom folder path - #6396
danielhanchen merged 29 commits into
Conversation
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces support for the 5.3.0 transformers tier (e.g., for Qwen3.5) and refactors the tier detection logic to check local config.json files before applying name heuristics. It also updates GPU estimation logging to handle expected 5.x-only config parsing failures gracefully and adds comprehensive unit tests. The reviewer suggests a high-severity improvement to recursively call get_transformers_tier(resolved) instead of _tier_from_name(resolved) when falling back to the resolved base model, ensuring the full tier detection logic is applied to the resolved model.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
…ckpoints A local safetensors folder whose config.json did not match the Gemma4 (510/550) architecture signals short-circuited get_transformers_tier() to "default" (transformers 4.57.x), never reaching the name-substring check that routes Qwen3.5 to the 5.3.0 sidecar. So a local Qwen3.5 checkpoint (model_type "qwen3_5", needs transformers >= 5.2.0) loaded with 4.57.x and failed with "does not support Qwen3.5". The same model as a remote HF id worked, because it has no local config.json to trigger the short-circuit. Detect the 5.3.0 tier from config.json (model_type "qwen3_5" / architecture Qwen3_5ForCausalLM) in the local-config branch, mirroring the existing Gemma4 510/550 handling. This is a positive config signal, so it fixes local Qwen3.5 without weakening the directory-name false-positive guard (a llama checkpoint under a "gemma-4-12b-*" parent still resolves to default). Adds tests for the config-based 530 detection and local-folder tier resolution.
…lies Expands the config.json-based tier detection to cover all known 5.3.0-tier model families (Qwen3 MoE, GLM-4.7-Flash, LFM2.5-VL) and adds a _name_or_path fallback so renamed local checkpoints with unrecognised model_type values still route correctly via the HF ID embedded in their config.json. - Expand _TRANSFORMERS_530_ARCHITECTURES / _MODEL_TYPES with verified entries from Qwen3MoeForCausalLM, Glm4MoeLiteForCausalLM, Lfm2VlForConditionalGeneration, and Qwen3_5ForConditionalGeneration (confirmed from local Qwen3.5-2B config.json) - Extract _tier_from_name() helper, deduplicating the fast-substring logic used by both the remote-path branch and the new config _name_or_path fallback - In the local-config branch: after architecture checks, resolve the tier from cfg._name_or_path / cfg.model_name before returning "default", preserving the existing directory-name false-positive guard - 79 tests passing
for more information, see https://pre-commit.ci
…emini suggestion)
3c83a33 to
043709b
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
unsloth/studio/backend/utils/transformers_version.py
Lines 586 to 589 in e294814
Add 5.3 config detection to the HF slow path
When a model is referenced by a Hub ID whose name does not contain one of the 5.3 substrings, this slow path still only checks 510/550 config and then falls through to tokenizer/default. That leaves the newly added qwen3_moe/qwen3_5/glm4_moe_lite config signals unused for renamed/private repos; it also affects local custom folders whose _name_or_path is resolved before activation, because activate_transformers_for_subprocess() calls _resolve_base_model() before get_transformers_tier(). In those cases a valid 5.3-only config can be classified as default and loaded under transformers 4.57.x.
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
… probes Using get_transformers_tier(resolved) on the _name_or_path fallback would trigger up to 3 network fetches (config.json + tokenizer_config.json, 10s each) for every ordinary checkpoint whose _name_or_path is a plain HF ID like meta-llama/Llama-3-8B. The fallback's purpose is name-based detection on the resolved HF ID, _tier_from_name covers all known cases without I/O.
Private or renamed HF repos whose model IDs lack a 5.3 substring were silently routed to the default tier. _check_config_needs_530 mirrors the existing 510/550 pattern: fetches config.json once, caches the result, and is called after the 550 check in the slow path. Includes 5 unit tests.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fdc074fc7e
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…ives When _name_or_path in config.json is an absolute path to the same checkpoint passed as a relative path, the textual resolved != model_name check passes and _tier_from_name would scan the directory path for substrings. Split the fallback: local directories recurse into get_transformers_tier (config check, no network I/O); HF Hub IDs use _tier_from_name (name-based, no network).
for more information, see https://pre-commit.ci
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 29952331f4
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
- _norm_separators(): collapse _ . whitespace to - so underscore/dot model ID variants (Qwen3_5, Qwen3_Next) match the canonical substring list - _tier_from_name(): apply norm to both name and each substring so aliases resolve without duplicating the substring lists - _resolve_base_model(): try model_name then _name_or_path separately so a self-referential Unsloth model_name doesn't hide the useful HF ID in _name_or_path - Gate get_base_model_from_lora on adapter_cfg_path.is_file() to avoid eagerly importing transformers before the sidecar venv is on sys.path - 17 new tests covering _norm_separators, separator-insensitive _tier_from_name, and the model_name/_name_or_path fallback
for more information, see https://pre-commit.ci
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: da225a669a
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
activate_transformers_for_subprocess and ensure_transformers_version were pre-resolving all local checkpoints via _resolve_base_model before calling get_transformers_tier. After the model_name/_name_or_path fix, a full checkpoint with a private/offline _name_or_path and no tier substring would resolve to that HF ID, which can't be probed, bypassing the local config.json model_type check entirely. Gate pre-resolution on adapter_config.json so full checkpoints go straight to get_transformers_tier, which reads config.json directly. LoRA adapters still pre-resolve as before.
for more information, see https://pre-commit.ci
|
tested again, LGTM after CI finishes 
|
…positives - Add Qwen3.5 MoE (qwen3_5_moe / Qwen3_5MoeForConditionalGeneration) and Qwen3-Next to the 5.3.0 config sets, so renamed local checkpoints route to the sidecar instead of default transformers - Let a 510/550 name match override a 530 config match, so Qwen3.6 (which reuses qwen3_5 / qwen3_5_moe config ids) still routes to the 5.5.0 sidecar - Stop normalizing version dots to hyphens so size names like Qwen3-5B and Qwen3-6B are not promoted to a 5.x sidecar; underscore aliases still match - Skip name matching for resolved values that look like stale local paths
for more information, see https://pre-commit.ci
|
Pushed a follow-up commit (cf72db5) on top of this to close a few gaps I hit while testing the tier detection against the live Hub configs. 1. Qwen3.5 MoE and Qwen3-Next were missing from the 5.3.0 config sets. The flagship Qwen3.5 MoE checkpoints (Qwen3.5-35B-A3B, Qwen3.5-122B-A10B) use 2. Qwen3.6 precedence. Qwen3.6 reuses the Qwen3.5 config ids (Qwen3.6-27B is 3. Dot-version false positives. 4. Stale Added regression tests for each; the suite is at 109 passing ( One thing I left out on purpose: I did not add |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 99b8d7f0bf
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
…ath-hint guard - adapter_model-only LoRA: add import-light _is_lora_adapter_dir/_has_adapter_weights and gate activation/export pre-resolve on them, so LoRA dirs with adapter_model*.safetensors but no adapter_config.json still resolve to their base model (via _resolve_base_model's new unsloth_<model>_<ts> directory-name parse) instead of tiering off the adapter folder. - 530 override: only treat a resolved value as a name hint when it is a real Hub id; a stale/renamed local path in model_name/_name_or_path can no longer flip a correct 530 config to 550. Current folder basename still allowed. Added 7 regression tests; suite at 116 passing.
for more information, see https://pre-commit.ci
|
Tested the latest commit end to end on Windows: a Qwen3.5-2B checkpoint loaded Before: After: Also pushed fixes for the two open codex P2s on top of Daniel's commit:
Added 7 regression tests, suite is at 116 passing. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1bae385d44
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| hint_src = ( | ||
| base | ||
| if (base != model_name and _looks_like_hf_id(base)) | ||
| else Path(model_name).name |
There was a problem hiding this comment.
When a local checkpoint's config positively matches the 530 set but has no usable HF ID, this falls back to the arbitrary folder basename and lets any higher-tier substring override the config. For example, a Qwen3.5 checkpoint saved in a custom directory named gemma-4-12b-experiment will select tier 510 even though its model_type/architecture says qwen3_5, defeating the local-config guard this change adds. The override should be limited to the specific Qwen3.6 signal or a trusted HF ID, rather than all 510/550 name matches from the local folder name.
Useful? React with 👍 / 👎.
|
|
||
| # Architecture classes / model_type values that require transformers 5.3.0. | ||
| # Checked via config.json (local or HuggingFace). | ||
| _TRANSFORMERS_530_ARCHITECTURES: set[str] = { |
There was a problem hiding this comment.
The new 530 config allow-list still omits the actual Ministral-3 identifiers: the upstream mistralai/Ministral-3-8B-Instruct-2512 config uses architectures: ["Mistral3ForConditionalGeneration"] and model_type: "mistral3". A locally saved Ministral-3 checkpoint in a custom folder without a useful _name_or_path therefore enters the local-config path, misses _config_needs_530, and returns default, so the worker won't activate the 5.3 sidecar for one of the families this module explicitly lists as 5.3-only.
Useful? React with 👍 / 👎.
|
@codex review |
|
/gemini review |
There was a problem hiding this comment.
Code Review
This pull request introduces support for a new transformers tier ('530') to accommodate models like Qwen3.5, Qwen3 MoE, GLM-4.7, and LFM2.5-VL. It refactors the tier detection and base model resolution logic in utils/transformers_version.py to be more robust, improves LoRA adapter detection for weight-only directories, and updates hardware configuration logging to gracefully handle sidecar-dependent models. The review feedback highlights several opportunities to enhance robustness and performance, such as leveraging the cached _load_config_json helper to avoid redundant disk I/O, defensively validating types and resolving paths during base model resolution, ensuring _looks_like_hf_id handles empty strings, and expanding exception handling in _is_lora_adapter_dir to prevent potential permission or OS-level crashes.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
| try: | ||
| with open(config_json_path) as f: | ||
| cfg = json.load(f) | ||
| # Unsloth writes "model_name"; HF writes "_name_or_path" | ||
| base = cfg.get("model_name") or cfg.get("_name_or_path") | ||
| if base and base != str(local_path): | ||
| logger.info( | ||
| "Resolved checkpoint '%s' → base model '%s' (via config.json)", | ||
| model_name, | ||
| base, | ||
| ) | ||
| return base | ||
| # Unsloth writes "model_name"; HF writes "_name_or_path". Try both: | ||
| # if "model_name" is self-referential (equals the local path), the | ||
| # useful base id may still live in "_name_or_path". | ||
| for _key in ("model_name", "_name_or_path"): | ||
| base = cfg.get(_key) | ||
| if base and base != str(local_path): | ||
| logger.info( | ||
| "Resolved checkpoint '%s' → base model '%s' (via config.json)", | ||
| model_name, | ||
| base, | ||
| ) | ||
| return base |
There was a problem hiding this comment.
To avoid redundant synchronous disk I/O and leverage the existing _config_json_cache, use the _load_config_json helper instead of manually opening and parsing config.json again.
| try: | |
| with open(config_json_path) as f: | |
| cfg = json.load(f) | |
| # Unsloth writes "model_name"; HF writes "_name_or_path" | |
| base = cfg.get("model_name") or cfg.get("_name_or_path") | |
| if base and base != str(local_path): | |
| logger.info( | |
| "Resolved checkpoint '%s' → base model '%s' (via config.json)", | |
| model_name, | |
| base, | |
| ) | |
| return base | |
| # Unsloth writes "model_name"; HF writes "_name_or_path". Try both: | |
| # if "model_name" is self-referential (equals the local path), the | |
| # useful base id may still live in "_name_or_path". | |
| for _key in ("model_name", "_name_or_path"): | |
| base = cfg.get(_key) | |
| if base and base != str(local_path): | |
| logger.info( | |
| "Resolved checkpoint '%s' → base model '%s' (via config.json)", | |
| model_name, | |
| base, | |
| ) | |
| return base | |
| try: | |
| cfg = _load_config_json(model_name) | |
| if cfg is not None: | |
| # Unsloth writes "model_name"; HF writes "_name_or_path". Try both: | |
| # if "model_name" is self-referential (equals the local path), the | |
| # useful base id may still live in "_name_or_path". | |
| for _key in ("model_name", "_name_or_path"): | |
| base = cfg.get(_key) | |
| if base and base != str(local_path): | |
| logger.info( | |
| "Resolved checkpoint '%s' → base model '%s' (via config.json)", | |
| model_name, | |
| base, | |
| ) | |
| return base |
References
- Centralize recurring or complex logical checks into a single helper function and reuse it across the codebase to ensure consistency and simplify maintenance.
| base = cfg.get(_key) | ||
| if base and base != str(local_path): |
There was a problem hiding this comment.
Defensively verify that base is a string (isinstance(base, str)) before comparing and returning it. If a malformed or custom config.json contains an unexpected type (like a list or dict) for model_name or _name_or_path, it could cause a TypeError crash later when passed to Path() or other string-based operations.
| base = cfg.get(_key) | |
| if base and base != str(local_path): | |
| base = cfg.get(_key) | |
| if base and isinstance(base, str) and base != str(local_path): |
| # useful base id may still live in "_name_or_path". | ||
| for _key in ("model_name", "_name_or_path"): | ||
| base = cfg.get(_key) | ||
| if base and base != str(local_path): |
There was a problem hiding this comment.
When base is a local path, comparing it directly as a string with str(local_path) can fail to detect that they point to the same directory if one is relative and the other is absolute. To make this comparison robust and avoid redundant recursion, resolve both paths when base is a local path.
| if base and base != str(local_path): | |
| if base and isinstance(base, str): | |
| try: | |
| base_path = Path(base) | |
| if base_path.exists() and base_path.resolve() == local_path.resolve(): | |
| continue | |
| except Exception: | |
| pass | |
| if base != str(local_path): |
| def _looks_like_hf_id(value: str) -> bool: | ||
| """True if *value* looks like a Hub id (``org/name``) rather than a local | ||
| filesystem path, so a stale/renamed checkpoint path isn't name-matched.""" | ||
| if os.path.isabs(value) or value.startswith((".", "~")) or "\\" in value: | ||
| return False | ||
| return value.count("/") <= 1 |
There was a problem hiding this comment.
Ensure that value is not empty or whitespace-only before performing the checks. An empty string would currently return True as a valid Hugging Face ID.
| def _looks_like_hf_id(value: str) -> bool: | |
| """True if *value* looks like a Hub id (``org/name``) rather than a local | |
| filesystem path, so a stale/renamed checkpoint path isn't name-matched.""" | |
| if os.path.isabs(value) or value.startswith((".", "~")) or "\\" in value: | |
| return False | |
| return value.count("/") <= 1 | |
| def _looks_like_hf_id(value: str) -> bool: | |
| """True if *value* looks like a Hub id (``org/name``) rather than a local | |
| filesystem path, so a stale/renamed checkpoint path isn't name-matched.""" | |
| if not value or not value.strip(): | |
| return False | |
| if os.path.isabs(value) or value.startswith((".", "~")) or "\\" in value: | |
| return False | |
| return value.count("/") <= 1 |
| def _is_lora_adapter_dir(path: Path) -> bool: | ||
| """True if *path* is a local LoRA adapter directory. | ||
|
|
||
| Mirrors ``utils.models._looks_like_lora_adapter`` but stays import-light so it | ||
| can run during subprocess activation without dragging in transformers. Detects | ||
| both ``adapter_config.json`` adapters and adapter_model-only LoRAs (weights | ||
| present, config absent) that a config-only check would miss. | ||
| """ | ||
| try: | ||
| if not path.is_dir(): | ||
| return False | ||
| except OSError: | ||
| return False | ||
| return (path / "adapter_config.json").is_file() or _has_adapter_weights(path) |
There was a problem hiding this comment.
Wrap the entire check in the try-except block to prevent potential OSError or PermissionError crashes when calling is_file() or checking adapter weights on directories with restricted permissions.
| def _is_lora_adapter_dir(path: Path) -> bool: | |
| """True if *path* is a local LoRA adapter directory. | |
| Mirrors ``utils.models._looks_like_lora_adapter`` but stays import-light so it | |
| can run during subprocess activation without dragging in transformers. Detects | |
| both ``adapter_config.json`` adapters and adapter_model-only LoRAs (weights | |
| present, config absent) that a config-only check would miss. | |
| """ | |
| try: | |
| if not path.is_dir(): | |
| return False | |
| except OSError: | |
| return False | |
| return (path / "adapter_config.json").is_file() or _has_adapter_weights(path) | |
| def _is_lora_adapter_dir(path: Path) -> bool: | |
| """True if *path* is a local LoRA adapter directory. | |
| Mirrors ``utils.models._looks_like_lora_adapter`` but stays import-light so it | |
| can run during subprocess activation without dragging in transformers. Detects | |
| both ``adapter_config.json`` adapters and adapter_model-only LoRAs (weights | |
| present, config absent) that a config-only check would miss. | |
| """ | |
| try: | |
| if not path.is_dir(): | |
| return False | |
| return (path / "adapter_config.json").is_file() or _has_adapter_weights(path) | |
| except OSError: | |
| return False |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1bae385d44
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if os.path.isabs(value) or value.startswith((".", "~")) or "\\" in value: | ||
| return False | ||
| return value.count("/") <= 1 |
There was a problem hiding this comment.
In the local-config fallback, _resolve_base_model() can return _name_or_path exactly as saved by Transformers; relative local values such as runs/qwen3.5-source (or a basename) are not absolute, do not start with ./~, and contain at most one slash, so this predicate treats them as Hub IDs and the fallback then applies _tier_from_name. A plain local Llama checkpoint with "_name_or_path": "runs/qwen3.5-source" is therefore routed to the 5.3 sidecar even though the config itself has no 5.x architecture, reintroducing the path-substring false positive the local config guard is meant to avoid.
Useful? React with 👍 / 👎.
| if _check_config_needs_550(model_name): | ||
| logger.info("Transformers tier 550 selected for %s (config.json check)", model_name) | ||
| return "550" | ||
| if _check_config_needs_530(model_name): |
There was a problem hiding this comment.
For a renamed/private Hub repo whose model ID lacks qwen3.6 but whose fetched config has model_type: qwen3_5 and _name_or_path: "Qwen/Qwen3.6-...", this slow-path check returns 530 immediately. The local-config branch above has an explicit Qwen3.6 override because those configs reuse Qwen3.5 IDs but require the 5.5 sidecar; the remote path should apply the same _name_or_path hint before selecting 530, otherwise these renamed Qwen3.6 repos load under transformers 5.3.0.
Useful? React with 👍 / 👎.
| _TRANSFORMERS_530_MODEL_TYPES: set[str] = { | ||
| "qwen3_5", | ||
| "qwen3_5_moe", | ||
| "qwen3_moe", | ||
| "qwen3_next", | ||
| "glm4_moe_lite", | ||
| "lfm2_vl", | ||
| } |
There was a problem hiding this comment.
The Qwen3.5 text-only configs use model types such as qwen3_5_text (and the MoE text config uses qwen3_5_moe_text), but this model-type set only includes the multimodal wrapper IDs. If a local fine-tuned/custom folder has one of those text config model_type values and its architectures entry is missing or stripped, _config_needs_530 returns false and the local-config path can fall back to default transformers, which cannot parse these Qwen3.5 text configs.
Useful? React with 👍 / 👎.
- Add Qwen3.5 text-tower model types (qwen3_5_text / qwen3_5_moe_text) to the 5.3.0 config set so text-only configs with stripped architectures still route to the sidecar - Apply the Qwen3.6 name override on the remote slow path too, so a renamed or private repo whose config reuses qwen3_5 ids but names Qwen3.6 in _name_or_path selects 5.5.0 instead of 5.3.0 - Treat an existing local path (or empty value) as a path, not a Hub id, in _looks_like_hf_id so a real local checkpoint folder is not name matched - Guard _resolve_base_model against non-string config values and compare paths by realpath so relative or absolute self references resolve correctly - Keep the LoRA adapter is_file check inside the OSError guard
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Pushed bde4f3f addressing the latest codex and gemini comments. Qwen3.5 text towers (codex P2). Added Remote Qwen3.6 override (codex P2). The slow config path now applies the same Relative path guard (codex P2).
On the gemini suggestion to route Suite is at 123 passing. |
|
@codex review |
|
Codex Review: Didn't find any major issues. Delightful! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
- _config_matches_tier no longer raises TypeError when a malformed config.json carries a non-string model_type (e.g. a list) or non-list architectures; it fails open to no-match - guard the model_name-derived is_file/is_dir probes with _safe_is_file / _safe_is_dir so a pathological or over-long path (e.g. a Windows long path) fails open to the default tier instead of raising OSError No routing changes for any valid model; purely defensive. Verified by a cross-platform simulation (POSIX + NT path semantics) and a before/after tier matrix that is unchanged for all previously supported models.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Ran an isolated cross-platform simulation of the tier detection (uv venv, POSIX and Windows path semantics via Before/after, across 32 representative cases: 12 previously-broken local custom-folder loads now route to the correct sidecar (Qwen3.5 dense/MoE/text, Qwen3-Next, Qwen3 MoE, GLM-4.7-Flash, LFM2.5-VL, Qwen3.6, and renamed folders), 20 previously-working cases are unchanged, and 0 mismatches. Mistral-Small-3 (which shares the While simulating edge cases I found and fixed two robustness gaps (810d3aa):
Both are purely defensive: the before/after tier matrix is identical for every valid model. Suite is at 128 passing. |
Shorten/remove over-long comments and docstrings, mainly on internal helpers, without changing behavior. Verified code-only via comment_tools.py check; suite unchanged at 128 passing.
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
1 similar comment
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Reconcile tier detection with main's NemotronH 5.10 routing (unslothai#6541) and SSM kernel auto-install (unslothai#6535). The production transformers_version.py auto-merged cleanly; resolve the test import block to the union of both symbol sets and update the adapter path-name recheck test to build a real adapter dir, so the resolved base drives the tier (path name never upgrades it).
013a108 to
2d50693
Compare
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
3 participants
Fixes #6395
What
get_transformers_tier()short-circuits to "default" (transformers 4.57.x)when a local
config.jsonis found but its architecture isn't in the knownsidecar sets, bypassing name-based detection entirely. This misroutes any
5.3.0-tier model loaded from a custom folder path.
Changes
_TRANSFORMERS_530_ARCHITECTURES/_MODEL_TYPESwith all confirmedfamilies: Qwen3.5, Qwen3 MoE, GLM-4.7-Flash, LFM2.5-VL
_resolve_base_model()(already extracts_name_or_path) and run substringdetection on the resolved HF ID, handles renamed folders without
directory-name false positives
hardware.pythat fires forsidecar-tier models whose config.json can't be parsed by default transformers
_tier_from_name()to deduplicate the substring logic used by boththe local-config fallback and the remote-path fast check