Skip to content

Define kv cache scaling factor as amax / 448 #790

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
cjluo-nv merged 4 commits into from
Jan 20, 2026
Merged

Define kv cache scaling factor as amax / 448 #790

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6019a6e
Define kv cache scaling factor as amax / 448
cjluo-nv Jan 16, 2026
852316f
Fix
cjluo-nv Jan 16, 2026
ba089a2
Merge branch 'main' into cjluo-nv-patch-3
cjluo-nv Jan 16, 2026
33b6dcb
Merge branch 'main' into cjluo-nv-patch-3
cjluo-nv Jan 20, 2026
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
19 changes: 3 additions & 16 deletions modelopt/torch/export/unified_export_hf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,9 +56,6 @@
set_expert_quantizer_amax,
)
from .model_config import (
KV_CACHE_FP8,
KV_CACHE_NVFP4,
KV_CACHE_NVFP4_AFFINE,
QUANTIZATION_FP8,
QUANTIZATION_FP8_PB_REAL,
QUANTIZATION_FP8_PC_PT,
Expand Down Expand Up @@ -647,19 +644,6 @@ def _export_transformers_checkpoint(

quant_config = get_quant_config(model, is_modelopt_qlora=is_modelopt_qlora)

kv_cache_max_bound = 0
kv_cache_format = quant_config["quantization"]["kv_cache_quant_algo"]

cache_bound_mapping = {
KV_CACHE_NVFP4: 6 * 448,
KV_CACHE_NVFP4_AFFINE: 6 * 448,
KV_CACHE_FP8: 448,
}

# Only update kv_cache_max_bound if a quantization is applied.
if kv_cache_format != QUANTIZATION_NONE:
kv_cache_max_bound = cache_bound_mapping.get(kv_cache_format)

# Process all quantized modules and export weights
_process_quantized_modules(model, dtype, is_modelopt_qlora)

Expand All @@ -669,6 +653,9 @@ def _export_transformers_checkpoint(
else:
quantized_state_dict = model.state_dict()

# We define kv cache scale as amax / 448 for both FP8 and NVFP4 KV cache quantization.
kv_cache_max_bound = 448
kv_cache_format = quant_config["quantization"]["kv_cache_quant_algo"]
quantized_state_dict = postprocess_state_dict(
quantized_state_dict, kv_cache_max_bound, kv_cache_format, is_modelopt_qlora
)
Expand Down