Skip to content

[None][fix] Align kv_scales with modelopt HF checkpoint #10745

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

Open
cjluo-nv wants to merge 5 commits into
base: main
Choose a base branch
from
Open

[None][fix] Align kv_scales with modelopt HF checkpoint #10745

Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
ba53212
Align kv_scales with modelopt HF checkpoint
cjluo-nv Jan 16, 2026
5e4138b
Fix scaling factor in copy_weight function
cjluo-nv Jan 16, 2026
291da15
Update
cjluo-nv Jan 20, 2026
b9bc035
Merge branch 'main' into cjluo-nv-patch-2
cjluo-nv Jan 21, 2026
6bdad6d
Merge branch 'main' into cjluo-nv-patch-2
cjluo-nv Jan 22, 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
12 changes: 2 additions & 10 deletions tensorrt_llm/_torch/modules/linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -577,14 +577,10 @@ def load_weights_fused_qkv_linear(self, module: Linear,
if os.environ.get("TRTLLM_LOAD_KV_SCALES", "0") == "1":
if len(k_scale) != 0:
assert len(v_scale) != 0
# The calibrated KV scales are amax / (6 * 448), but the requested KV scales are amax / 448,
# to avoid overflow when dequantizing NVFP4 in attention kernels.
copy_weight(
module.kv_scales,
torch.tensor(
[1.0, max(k_scale) * 6.0,
max(v_scale) * 6.0],
dtype=torch.float32))
[1.0, max(k_scale), max(v_scale)], dtype=torch.float32))
module.inv_kv_scales.data = 1.0 / module.kv_scales

def load_weights_fused_gate_up_linear(self, module: Linear,
Expand Down Expand Up @@ -1096,14 +1092,10 @@ def load_weights_fused_qkv_linear(self, module: Linear,
if os.environ.get("TRTLLM_LOAD_KV_SCALES", "0") == "1":
if len(k_scale) != 0:
assert len(v_scale) != 0
# The calibrated KV scales are amax / (6 * 448), but the requested KV scales are amax / 448,
# to avoid overflow when dequantizing NVFP4 in attention kernels using FP8 math.
copy_weight(
module.kv_scales,
torch.tensor(
[1.0, max(k_scale) * 6.0,
max(v_scale) * 6.0],
dtype=torch.float32))
[1.0, max(k_scale), max(v_scale)], dtype=torch.float32))
module.inv_kv_scales.data = 1.0 / module.kv_scales

def load_weights_fused_gate_up_linear(self, module: Linear,
Expand Down