Skip to content

feat: add support for Modelopt fp8_pb_wo quantization scheme #6106

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
merged 4 commits into from
Jul 18, 2025
+13 −4
Merged

feat: add support for Modelopt fp8_pb_wo quantization scheme #6106

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
327d42d
feat: add support for Modelopt fp8_pb_wo quantization scheme
achartier Jul 16, 2025
cef7781
Merge branch 'main' into fp8_pb_wo
symphonylyh Jul 17, 2025
58c5594
address review
achartier Jul 18, 2025
aa41757
Merge branch 'main' into fp8_pb_wo
achartier Jul 18, 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
3 changes: 3 additions & 0 deletions tensorrt_llm/_torch/model_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -202,6 +202,9 @@ def from_pretrained(cls,
json_quant_configs = quant_config_dict['quantization']

quant_config.quant_algo = json_quant_configs.get('quant_algo', None)
# fp8_pb_wo from modelopt is the same as FP8_BLOCK_SCALES
if quant_config.quant_algo == "fp8_pb_wo":
quant_config.quant_algo = 'FP8_BLOCK_SCALES'
quant_config.kv_cache_quant_algo = json_quant_configs.get(
'kv_cache_quant_algo', None)
quant_config.group_size = json_quant_configs.get('group_size', None)
Expand Down
8 changes: 5 additions & 3 deletions tensorrt_llm/_torch/modules/linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -562,7 +562,8 @@ def load_weights_vanilla(self, module: Linear, weights: List[Dict]) -> None:

scale_name = self._get_scale_name(weights)
weight_scale = load_weight_shard(weights[0][scale_name], module.tp_size,
module.tp_rank, module.tp_mode)
module.tp_rank,
module.tp_mode).squeeze()
copy_weight(module.weight_scale, weight_scale)
if "input_scale" in weights[0]:
copy_weight(module.input_scale, weights[0]["input_scale"])
Expand All @@ -582,7 +583,8 @@ def load_weights_fused_qkv_linear(self, module: Linear,
module.tp_rank, module.tp_mode)
v_scale = load_weight_shard(weights[2][scale_name], module.tp_size,
module.tp_rank, module.tp_mode)
fused_fp8_block_scale = torch.cat((q_scale, k_scale, v_scale))
fused_fp8_block_scale = torch.cat((q_scale, k_scale, v_scale)).squeeze()

copy_weight(module.weight_scale, fused_fp8_block_scale)

def load_weights_fused_gate_up_linear(self, module: Linear,
Expand All @@ -597,7 +599,7 @@ def load_weights_fused_gate_up_linear(self, module: Linear,
module.tp_rank, module.tp_mode)
right_scale = load_weight_shard(weights[1][scale_name], module.tp_size,
module.tp_rank, module.tp_mode)
fused_scale = torch.cat([left_scale, right_scale], dim=0)
fused_scale = torch.cat([left_scale, right_scale], dim=0).squeeze()
copy_weight(module.weight_scale, fused_scale)


Expand Down
6 changes: 5 additions & 1 deletion tensorrt_llm/llmapi/llm_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -362,7 +362,11 @@ def _update_from_hf_quant_config(self) -> bool:

hf_quant_algo = hf_quant_config.pop("quant_algo", None)
if hf_quant_algo is not None:
hf_quant_algo = QuantAlgo(hf_quant_algo)
# fp8_pb_wo from modelopt is the same as fp8_block_scales
if hf_quant_algo == "fp8_pb_wo":
hf_quant_algo = QuantAlgo.FP8_BLOCK_SCALES
else:
hf_quant_algo = QuantAlgo(hf_quant_algo)
if quant_config.quant_algo is None:
logger.info(
f"Setting quant_algo={hf_quant_algo} form HF quant config."
Expand Down