Revert "docs: add safetensors mirror of Qwen/Qwen3.6-27B"

This reverts commit b4203787cd51fa6cb79ab320f3702f0253303557.

CHANGELOG.md

@@ -7,42 +7,6 @@ and documentation**, not the underlying base model.
 
 ## [Unreleased]
 
-### Removed (transformers config)
-- **Dropped `config.json`** (`5302d10`) to suppress HF's tag
-  auto-detector surfacing `qwen3_5` in the repo header — the
-  detector reads `architectures` from `config.json` and the
-  surfaced tag was obscuring this card's positioning.
-  Consequence: `AutoModelForCausalLM.from_pretrained(
-  "FoolDev/Thanatos-27B")` no longer works on its own.
-  `examples/transformers_quickstart.py` now pulls `AutoConfig`
-  from upstream `Qwen/Qwen3.6-27B` (byte-identical tensors,
-  so no behavioural difference) and weights + tokenizer +
-  chat template from this repo. README's "What's here"
-  table and transformers paragraph updated to match.
-
-### Added (safetensors mirror)
-- **Mirrored Qwen/Qwen3.6-27B's transformers-loadable safetensors
-  set into this repo.** 15 sharded `.safetensors` files (~58 GB
-  total) + `model.safetensors.index.json` + tokenizer files
-  (`tokenizer.json`, `tokenizer_config.json`, `vocab.json`,
-  `merges.txt`) + configs (`configuration.json`,
-  `generation_config.json`, `preprocessor_config.json`,
-  `video_preprocessor_config.json`) + `chat_template.jinja`.
-  (`config.json` was initially mirrored too, then dropped — see
-  "Removed (transformers config)" above.) Tensor data
-  byte-identical to upstream; the mirror saves a second
-  `hf download` for users who want both GGUF + safetensors in
-  one place. `.gitignore` was updated separately (commit
-  `0c5bee4`) to whitelist the Qwen sharded naming pattern before
-  the upload's preupload check ran (HF reads the destination
-  repo's `.gitignore` to decide `shouldIgnore` per file).
-- `examples/transformers_quickstart.py` defaults `MODEL_ID`
-  to `FoolDev/Thanatos-27B` (weights + tokenizer + chat
-  template) with `CONFIG_ID="Qwen/Qwen3.6-27B"` for the
-  architecture config — fresh users still need only this
-  repo as the entry point, with one auxiliary HF Hub pull
-  for `config.json` that transformers handles transparently.
-
 ### Changed (5th round trip — qwen36 → qwen35, retested next-day)
 - **Bundle re-stamped `general.architecture: 'qwen36'` → `'qwen35'`**
   in `hf upload` commit `e03e10e` (HF), 2026-05-20 midday — 8

README.md

@@ -125,9 +125,6 @@ The 27B is **dense**: every parameter participates in every forward pass. It's s
 | `scripts/verify_arch.py` | Cross-checks the README "Architecture" forward-pass bullets (layer count, head counts, hidden / FFN dims, RoPE factor, SSM dims, vocab, context) against the actual GGUF metadata keys. Run as `make verify-arch`. Handles both `qwen35`- and `qwen36`-stamped bundles; exit non-zero if any value mismatches. Not part of `make check` because it loads the 17 GB GGUF (LFS smudge required); run on demand. |
 | `scripts/install-hooks.sh` | Installs `check.sh` as a git pre-commit hook |
 | `Makefile` | Convenience wrapper — `make help` lists targets |
-| `model-*-of-00015.safetensors` (15 files, ~58 GB) + `model.safetensors.index.json` | Transformers-loadable safetensors mirror of `Qwen/Qwen3.6-27B`. Byte-identical to upstream. |
-| `configuration.json`, `generation_config.json`, `preprocessor_config.json`, `video_preprocessor_config.json`, `chat_template.jinja` | Processor + chat-template configs mirrored from upstream. **`config.json` is intentionally not in this repo** — HF's tag auto-detector reads `architectures` from it and surfaces `qwen3_5` in the repo header, which obscures this repo's positioning. Transformers users: pull `config.json` from [`Qwen/Qwen3.6-27B`](https://huggingface.co/Qwen/Qwen3.6-27B) (see Transformers note below). |
-| `tokenizer.json`, `tokenizer_config.json`, `vocab.json`, `merges.txt` | Tokenizer files mirrored from upstream. |
 | `LICENSE`, `CITATION.cff` | Apache-2.0 license and citation metadata |
 | `CHANGELOG.md` | Versioned tooling/docs changes |
 | `README.md` | This file |
@@ -141,28 +138,7 @@ local-build path applies this repo's `Modelfile`; the `hf.co/...`
 path applies the root-level `template`, `system`, and `params`
 files (kept in sync with the `Modelfile`).
 
-The transformers safetensors set is mirrored in this repo
-(15 sharded `.safetensors` files + index + tokenizer +
-chat template), byte-identical to upstream
-[`Qwen/Qwen3.6-27B`](https://huggingface.co/Qwen/Qwen3.6-27B).
-**`config.json` is not bundled here** — HF auto-detects model
-architecture from it and surfaces a `qwen3_5` repo-level tag
-that obscures this card. To load via transformers, either:
-
-```python
-# A. Use upstream as the config/architecture source, this repo for weights:
-from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
-cfg = AutoConfig.from_pretrained("Qwen/Qwen3.6-27B", trust_remote_code=True)
-tok = AutoTokenizer.from_pretrained("FoolDev/Thanatos-27B", trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(
-    "FoolDev/Thanatos-27B", config=cfg, trust_remote_code=True,
-)
-
-# B. Or just load upstream directly — tensors are byte-identical:
-model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-27B", trust_remote_code=True)
-```
-
-`examples/transformers_quickstart.py` uses path A.
+If you want the safetensors for `transformers`, fetch them from [`Qwen/Qwen3.6-27B`](https://huggingface.co/Qwen/Qwen3.6-27B).
 
 ## Architecture
 
@@ -581,7 +557,7 @@ python examples/ollama_chat.py        # section 3 runs a real round-trip
 
 | Model | Notes |
 |---|---|
-| [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) | Upstream base, safetensors (this repo mirrors them) |
+| [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) | Upstream base, safetensors |
 | [unsloth/Qwen3.6-27B-GGUF](https://huggingface.co/unsloth/Qwen3.6-27B-GGUF) | Recommended GGUF source |
 | [FoolDev/Janus-35B](https://huggingface.co/FoolDev/Janus-35B) | 35B-A3B MoE sibling. More capacity, more memory pressure. |
 | [Crownelius/Crow-9B-HERETIC-4.6](https://huggingface.co/Crownelius/Crow-9B-HERETIC-4.6) | 9B starter model when 27B/35B is too heavy |

examples/README.md

@@ -5,7 +5,7 @@ Four minimal entry points. Pick the one that matches how you run models.
 | File | Backend | When to use |
 |---|---|---|
 | `ollama_chat.py` | Ollama HTTP API | You already have `ollama serve` running and the `thanatos-27b` model created from the project `Modelfile`. **Text + tool calling** — vision via Ollama is broken upstream for this arch. |
-| `transformers_quickstart.py` | Hugging Face Transformers | You want to run the safetensors on GPU (this repo mirrors them from `Qwen/Qwen3.6-27B`), optionally in 4-bit via bitsandbytes. |
+| `transformers_quickstart.py` | Hugging Face Transformers | You want to run the upstream safetensors (`Qwen/Qwen3.6-27B`) on GPU, optionally in 4-bit via bitsandbytes. |
 | `llama_cpp_quickstart.py` | llama-cpp-python | You want to invoke a local GGUF directly without a daemon (CI, batch jobs, scripts). Text only. |
 | `llama_cpp_vision.py` | llama-cpp-python + mmproj | **Image input.** Loads a text GGUF + `mmproj-F16.gguf` and answers questions about an image. The only working vision path right now. |
 

examples/transformers_quickstart.py

@@ -2,17 +2,11 @@
 """
 Thanatos-27B — Hugging Face Transformers quickstart.
 
-Loads the Qwen 3.6 27B safetensors from this repo (a byte-identical
-mirror of Qwen/Qwen3.6-27B's transformers set) and runs a single
-chat turn using its embedded chat template. Applies the same
-Thanatos system prompt the Modelfile / bridge `system` file uses.
-
-`config.json` is intentionally not in this repo (it makes HF's tag
-auto-detector surface a `qwen3_5` repo-level tag), so we source the
-architecture config from upstream `Qwen/Qwen3.6-27B` and only pull
-weights + tokenizer + chat template from this repo. Tensor data is
-byte-identical, so the result is the same model. Set
-`MODEL_ID = "Qwen/Qwen3.6-27B"` to bypass this repo entirely.
+Loads the upstream Qwen 3.6 27B safetensors directly and runs a single
+chat turn using its embedded chat template. Thanatos-27B is a *wrapper*
+around that base, so for the transformers route there is nothing to
+download from this repo — point at Qwen/Qwen3.6-27B and apply the same
+system prompt the Modelfile uses.
 
 Requirements:
     pip install --upgrade "transformers>=4.45" accelerate sentencepiece bitsandbytes
@@ -34,7 +28,7 @@ import sys
 
 try:
     import torch
-    from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+    from transformers import AutoModelForCausalLM, AutoTokenizer
 except ImportError as e:  # pragma: no cover
     sys.exit(
         f"Missing dependency: {e.name}. Install with:\n"
@@ -42,8 +36,7 @@ except ImportError as e:  # pragma: no cover
     )
 
 
-MODEL_ID = "FoolDev/Thanatos-27B"
-CONFIG_ID = "Qwen/Qwen3.6-27B"  # source of config.json (not bundled in MODEL_ID — see module docstring)
+MODEL_ID = "Qwen/Qwen3.6-27B"
 
 THANATOS_SYSTEM = (
     "You are Thanatos, a precise and capable assistant for reasoning, writing, "
@@ -75,11 +68,8 @@ def load(use_4bit: bool):
         )
         kwargs.pop("torch_dtype", None)
 
-    cfg = AutoConfig.from_pretrained(CONFIG_ID, trust_remote_code=True)
     tok = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
-    model = AutoModelForCausalLM.from_pretrained(
-        MODEL_ID, config=cfg, trust_remote_code=True, **kwargs,
-    )
+    model = AutoModelForCausalLM.from_pretrained(MODEL_ID, trust_remote_code=True, **kwargs)
     return tok, model