[Feature][Humming] Humming feature integration (#56685)

Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Jinzhen Lin
2026-09-20 08:11:01 -07:00
committed by GitHub
co-authored by Codex mgoin
parent 27757dde02
commit 7d99c2c4fd
62 changed files with 2040 additions and 1086 deletions
@@ -4,11 +4,13 @@
import torch
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.utils.humming_utils import (
from vllm.model_executor.layers.quantization.utils.humming import (
apply_humming_linear,
get_humming_linear_compute_config,
prepare_humming_linear_layer_config,
quant_key_to_input_schema,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import kNvfp4Dynamic
from vllm.platforms import current_platform
from vllm.utils.import_utils import has_humming
@@ -40,9 +42,7 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
# Route through humming's compressed-tensors nvfp4 loader (same path as
# the MoE oracle); the native group_tensor schema mishandles a scalar
# global scale.
# Reuse the CT loader for packed FP4 weights and tensor scales.
quant_config = {
"quant_method": "compressed-tensors",
"format": "nvfp4-pack-quantized",
@@ -60,7 +60,16 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
layer.weight_global_scale = torch.nn.Parameter(
1.0 / layer.weight_global_scale, requires_grad=False
)
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
input_quant_key = None
if hasattr(layer, "input_global_scale"):
layer.input_scale_2 = torch.nn.Parameter(
layer.input_global_scale.reshape(-1), requires_grad=False
)
input_quant_key = kNvfp4Dynamic
input_schema = quant_key_to_input_schema(input_quant_key)
self.layer_config = prepare_humming_linear_layer_config(
layer, quant_config, input_schema=input_schema
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
@@ -73,6 +82,7 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
return apply_humming_linear(
layer,
x,
skip_bias_add=bias is None,
layer_config=self.layer_config,
compute_config=self.compute_config,
locks=self.locks,