vllm-project/tpu-inference

[Feature] Deprecate pre-quantized model loading via Qwix for the JAX/Flax path

Open

#1680 aperta il 10 feb 2026

Vedi su GitHub
 (2 commenti) (0 reazioni) (1 assegnatario)Python (205 fork)github user discovery
good first issue

Metriche repository

Star
 (348 star)
Metriche merge PR
 (Metriche PR in attesa)

Descrizione

🚀 The feature, motivation and pitch

Once #1623 is complete, we should fully deprecate using Qwix to load prequantized models (example)

More specifically, we need to do the followng:

  1. Deprecate the usage of apply_qwix_on_abstract_model, since we should only be applying Qwix "on-the-fly" (that is, on an unquantized model). We only apply need to Qwix on the abstract model in the case that the model is pre-quantized
  2. Deprecate all of the "default" quantization logic
  3. Convert Llama4 and GPT-OSS to avoid using Qwix to load the pre-quantized model

Alternatives

No response

Additional context

No response

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

Guida contributor