Clément Dumas
Clément Dumas
fyi, my nnterp renaming tests are passing, see: https://github.com/Butanium/nnterp/blob/0.4renamingdev/tests/test_model_renaming.py. I have on my todo to add some tests to nnsight directly but this might come only after the neurips deadline.
Maybe let's wait until the tests are implemented?
oh yeah that would be nice!
Acutally, replacing the last display by a print: ```py print(cache.model.transformer.h[0]) ``` prints the full cache ```py {'model.transformer.h.0': Cache.Entry(output=(tensor([[[ 6.7837e-01, -1.4596e+00, 8.3400e-01, -1.5560e+00, 3.3362e-01, -8.6388e-01, 7.9839e-02, -1.6176e+00, ...]]]),), inputs=None) , 'model.transformer.h.1':...
Script i used to find this: ```py #!/usr/bin/env python3 """ Script to find the minimum transformers version compatible with torch>=2.4.0 Tests versions in descending order until we hit a dependency...
bump @AdamBelfki3. Idk if it's an easy fix but that'd be great to have this fixed upstream for nnterp
@AdamBelfki3 thx, still some problem with path renaming: ```py from nnsight import LanguageModel m = LanguageModel("gpt2", rename={"transformer.h": "layers"}) with m.trace("Hi") as t: cache = t.cache(modules=[m.layers[0]]).save() print("Keys:", list(cache.keys())) print("Trying: cache.layers[0].output") print(cache.model.layers[0].output)...
Oh yes sorry Claude was being dumb
@AdamBelfki3 i updated my example which also crashes
@AdamBelfki3 sorry this is the minimal dict that crashes for me! ```py { "transformer": "model", "h": "layers", "model.layers": "layers", } ```