WanVideoTextEncodeCached.process() (nodes.py) writes/reads text-embedding tensors to a disk cache under text_embed_cache/, keyed only on the prompt string:
def get_cache_path(prompt):
cache_key = prompt.strip()
cache_hash = hashlib.sha256(cache_key.encode('utf-8')).hexdigest()
return os.path.join(cache_dir, f"{cache_hash}.pt")(nodes.py, get_cache_path, ~lines 156-159, called from get_cached_text_embeds/process around line 258)
process() also takes model_name, precision (fp32/bf16), and quantization (disabled/fp8_e4m3fn) as inputs — all of which affect the embedding tensor that gets produced — but none of them are part of the cache key. Practical effect: encode a prompt once with encoder A, then switch to encoder B (or just change precision/quantization) while reusing the same prompt text, and the node silently serves the embeddings computed with the old model/precision from disk instead of re-encoding. No error, no warning — and since the cache is on disk, this persists across ComfyUI restarts, not just within a session.
Suggested fix: fold model_name, precision, and quantization into the hashed key, e.g.
cache_key = f"{model_name}|{precision}|{quantization}|{prompt.strip()}"Existing cache files would simply go unused and get regenerated under new hashes — no migration needed.
Minor/secondary, same mechanism: the in-memory _extender_cache a few lines up (keyed on (orig_prompt, str(extender_args))) similarly doesn't include the device/precision args passed to the prompt-extender's generate() call. Lower severity since it's LLM-generated prompt text rather than embeddings, but flagging it here since it's the same pattern in the same function.
WanVideoTextEncodeCached.process()(nodes.py) writes/reads text-embedding tensors to a disk cache undertext_embed_cache/, keyed only on the prompt string:(nodes.py,
get_cache_path, ~lines 156-159, called fromget_cached_text_embeds/processaround line 258)process()also takesmodel_name,precision(fp32/bf16), andquantization(disabled/fp8_e4m3fn) as inputs — all of which affect the embedding tensor that gets produced — but none of them are part of the cache key. Practical effect: encode a prompt once with encoder A, then switch to encoder B (or just change precision/quantization) while reusing the same prompt text, and the node silently serves the embeddings computed with the old model/precision from disk instead of re-encoding. No error, no warning — and since the cache is on disk, this persists across ComfyUI restarts, not just within a session.Suggested fix: fold
model_name,precision, andquantizationinto the hashed key, e.g.Existing cache files would simply go unused and get regenerated under new hashes — no migration needed.
Minor/secondary, same mechanism: the in-memory
_extender_cachea few lines up (keyed on(orig_prompt, str(extender_args))) similarly doesn't include thedevice/precisionargs passed to the prompt-extender'sgenerate()call. Lower severity since it's LLM-generated prompt text rather than embeddings, but flagging it here since it's the same pattern in the same function.