Skip to content

fix: enabling block-by-block evaluation for granite-3.x-models #165

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
chichun-charlie-liu merged 3 commits into from
Jul 24, 2025
Merged

fix: enabling block-by-block evaluation for granite-3.x-models #165

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
bf53238
fix: Enabled eval_llm_1GPU for granite-3.x-models
bayo-ibm Jul 22, 2025
6c1da4c
fix: updated ways of getting obtaining logits_scaling
bayo-ibm Jul 23, 2025
3d5b342
fix: added comments to the logits scaling calculation in eval_llm_1GPU
bayo-ibm Jul 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion fms_mo/quant/ptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2140,14 +2140,22 @@ def get_blocks(model, model_type=None):
None,
"lm_head",
),
"granite": (
"granite_gptbigcode": (
"transformer.h",
"transformer.wte",
"transformer.wpe",
None,
"transformer.ln_f",
"lm_head",
),
"granite": (
"model.layers",
"model.embed_tokens",
"model.rotary_emb",
None,
"model.norm",
"lm_head",
),
"llama": (
"model.layers",
"model.embed_tokens",
Expand Down
1 change: 0 additions & 1 deletion fms_mo/run_quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -155,7 +155,6 @@ def run_gptq(model_args, data_args, opt_args, gptq_args):
v2_memory_device="cpu",
)


# Add custom model_type mapping to gptqmodel LUT so GPTQModel can recognize them.
for mtype, cls in custom_gptq_classes.items():
if mtype in MODEL_MAP:
Expand Down
5 changes: 2 additions & 3 deletions fms_mo/training_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,8 +139,8 @@ class OptArguments(TypeChecker):

quant_method: str = field(
metadata={
"choices": ["gptq", "gptqv2", "fp8", "dq"],
"help": "Quantization technique"
"choices": ["gptq", "gptqv2", "fp8", "dq"],
"help": "Quantization technique",
}
)
output_dir: str = field(
Expand Down Expand Up @@ -229,7 +229,6 @@ class GPTQArguments(TypeChecker):
cache_examples_on_gpu: bool = True



@dataclass
class FP8Arguments(TypeChecker):
"""Dataclass for FP8 related arguments that will be used by llm-compressor."""
Expand Down
7 changes: 7 additions & 0 deletions fms_mo/utils/eval_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,6 +98,10 @@ def eval_llm_1GPU(qcfg, model, test_dataset, pre_cache_func=None, **kwargs): #
logger.info("All blocks are computed for evaluation")

nlls = []

# Required by Granite-3.X (and potentially other models) to scale their logits
logits_scaling = getattr(model.config, "logits_scaling", 1)

# for i, data_mb in enumerate(dloader): #if using dloader.
for i in tqdm(range(qcfg["n_samples"]), desc="Final Evaluating..."):
hidden_states = qcfg["cached_input"][i].to(dev)
Expand All @@ -107,6 +111,9 @@ def eval_llm_1GPU(qcfg, model, test_dataset, pre_cache_func=None, **kwargs): #
lm_head.to(dev)
lm_logits = lm_head(hidden_states)

# Scaling of the lm_head outputs to obtain the correct logits
lm_logits /= logits_scaling

# Shift so that tokens < n predict n
shift_logits = lm_logits[:, :-1, :].contiguous().float()
shift_labels = test_dataset.input_ids[:, (i * seq_len) : ((i + 1) * seq_len)][
Expand Down
Loading