WIP: janice network issues

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

src/llmcompressor/modifiers/transform/quip/base.py

@@ -45,6 +45,7 @@ class QuIPModifier(Modifier):
     )
     randomize: bool = Field(default=False, exclude=True)
     learnable: bool = Field(default=False, exclude=True)
+    precision: 
     ignore: Union[str, List[str]] = Field(default="lm_head", exclude=True)
 
     # optional override for more fine-grained control

tests/llmcompressor/modifiers/transform/test_correctness.py

@@ -1,3 +1,4 @@
+import os
 import pytest
 import torch
 from transformers import AutoModelForCausalLM
@@ -8,16 +9,20 @@
 
 
 @requires_gpu
+# @pytest.mark.skipif(
+#     (not os.getenv("HF_TOKEN")),
+#     reason="Skipping tracing tests requiring gated model access",
+# )
 @pytest.mark.parametrize(
     "dtype,exp_mse",
     [
-        (torch.bfloat16, 1e-2),
-        (torch.float32, 1e-9),
+        (torch.bfloat16, 5e-3),
+        (torch.float32, 5e-11),
     ],
 )
 def test_apply_correctness(dtype, exp_mse):
     model = AutoModelForCausalLM.from_pretrained(
-        "meta-llama/Meta-Llama-3-8B-Instruct", device_map="cuda", torch_dtype=dtype
+        "meta-llama/Llama-3.2-1B-Instruct", device_map="cuda", torch_dtype=dtype
     )
     state = State(model=model)
     modifier = QuIPModifier(transform_type="random-hadamard")
@@ -32,4 +37,5 @@ def test_apply_correctness(dtype, exp_mse):
     with torch.no_grad():
         output = model(**input)
 
+    print(torch.nn.MSELoss()(output.logits, true_output.logits))
     assert torch.nn.MSELoss()(output.logits, true_output.logits) <= exp_mse