Hard-pin transformers==4.57.1 (was >=5.6,<5.8)

The earlier 5.6/5.7/5.8 fixes were chasing the wrong target. The real
break happens **across the 4.x → 5.x boundary**: transformers 5.x
moved `config.rope_theta` into a `rope_parameters` dict and never
populates the top-level attribute, even when supplied via kwargs. The
upstream `modeling_qwen3.py` we vendor (commit df86ca9) reads
`config.rope_theta` directly, so on any 5.x the rotary frequencies
silently default — base attention still runs, but with wrong positions,
the gen tower output collapses, and 50-step sampling produces a flat
mid-grey image with patch-grid texture (everything else — model load,
LoRA wrap, ckpt inject, even `think_mode` text generation — succeeds
without warning).

This was masked because the venv was originally on 4.57.1 (May 1 - May 6)
where v16c sampling worked. The publish-cleanup commit relaxed
`transformers>=5.6` and a subsequent `pip install -e .` (May 7) silently
upgraded the local venv to 5.8.0, which is when sampling broke.

Hard-pin all three requirements specs to the exact 4.57.1 we know works.
The earlier 5.x compatibility shims (rope_theta auto-promotion,
all_tied_weights_keys, compute_default_rope_parameters method) stay in
the snapshot files as no-ops on 4.57.1; they only kick in if a future
user is on 5.x. Same for the loader.py self-heal for transformers'
indirect-import dynamic-cache miss.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

pyproject.toml

@@ -32,10 +32,13 @@ classifiers = [
 ]
 dependencies = [
   "torch>=2.9",
-  # Upper bound: transformers 5.8 broke the upstream `modeling_qwen3.py`
-  # we pin (rope_theta moved into rope_parameters dict, all_tied_weights_keys
-  # appeared, RotaryEmbedding now needs `compute_default_rope_parameters`).
-  "transformers>=5.6,<5.8",
+  # Hard-pin: the upstream `modeling_qwen3.py` we vendor (commit df86ca9)
+  # was written against transformers 4.57.1 and reads `config.rope_theta`
+  # directly. transformers 5.x moved that into a `rope_parameters` dict,
+  # which silently breaks attention frequencies and produces noise images
+  # at sample time even when the LoRA + base load cleanly. Stay on 4.57.1
+  # until we either re-train or rewrite the modeling files.
+  "transformers==4.57.1",
   "safetensors>=0.7",
   "huggingface_hub>=0.26",
   "sentencepiece",

requirements.txt

@@ -3,11 +3,10 @@
 
 # core
 torch>=2.9
-# Upper bound: transformers 5.8 reshuffled rope params API + tied_weights
-# ABI in incompatible ways with the upstream `modeling_qwen3.py` we pin.
-# See loader.py for the indirect-import self-heal that survives upstream
-# transformers churn beyond what version-pinning catches.
-transformers>=5.6,<5.8
+# Hard-pin: the upstream modeling_qwen3.py (commit df86ca9) reads
+# `config.rope_theta` directly; transformers 5.x moved that into a
+# `rope_parameters` dict, silently breaking attention freqs at sample time.
+transformers==4.57.1
 safetensors>=0.7
 
 # HF model load (NEOChatModel uses trust_remote_code)

train_u1/requirements.txt

@@ -4,7 +4,7 @@
 
 # core
 torch>=2.9
-transformers>=5.6,<5.8       # 5.8 broke the pinned upstream modeling_qwen3.py
+transformers==4.57.1       # upstream modeling_qwen3.py reads config.rope_theta directly; 5.x dict-form silently breaks attention
 safetensors>=0.7
 
 # HF model load (NEOChatModel uses trust_remote_code)