[Quant][Kernel] Remove GPTQ Group/Dynamic Activation Ordering (#54809)

Signed-off-by: Roderick-Wu <roderickwu2003@gmail.com>
Signed-off-by: Roderick Wu <roderickwu2003@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Roderick Wu
2026-09-08 12:45:12 -07:00
committed by GitHub
co-authored by mgoin
parent 1a522b6949
commit 2c9d68f80b
99 changed files with 1182 additions and 3917 deletions
@@ -272,13 +272,13 @@ def _stub_marlin_ops(monkeypatch):
monkeypatch.setattr(
ops,
"gptq_marlin_repack",
lambda w, perm, size_k, size_n, num_bits, is_a_8bit=False: torch.zeros(
lambda w, size_k, size_n, num_bits, is_a_8bit=False: torch.zeros(
size_k // 16, size_n * 2, dtype=torch.int32
),
)
def _make_act_order_marlin_kernel():
def _make_marlin_kernel():
from vllm.model_executor.kernels.linear.mixed_precision.marlin import (
MarlinLinearKernel,
)
@@ -295,20 +295,17 @@ def _make_act_order_marlin_kernel():
act_type=torch.float16,
group_size=_MARLIN_GROUP_SIZE,
zero_points=False,
has_g_idx=True,
)
kernel.w_q_name = "qweight"
kernel.w_s_name = "scales"
kernel.w_zp_name = None
kernel.w_gidx_name = "g_idx"
return kernel
def _load_marlin_checkpoint_format_weights(layer, g_idx):
def _load_marlin_checkpoint_format_weights(layer):
from vllm.model_executor.parameter import (
GroupQuantScaleParameter,
PackedvLLMParameter,
RowvLLMParameter,
)
layer.qweight = PackedvLLMParameter(
@@ -327,54 +324,24 @@ def _load_marlin_checkpoint_format_weights(layer, g_idx):
output_dim=1,
weight_loader=default_weight_loader,
)
layer.g_idx = RowvLLMParameter(
data=g_idx.clone(),
input_dim=0,
weight_loader=default_weight_loader,
)
def _random_g_idx(generator):
return torch.randint(
0,
_MARLIN_SIZE_K // _MARLIN_GROUP_SIZE,
(_MARLIN_SIZE_K,),
dtype=torch.int32,
generator=generator,
)
def test_marlin_post_load_preserves_runtime_tensor_addresses(monkeypatch, dist_init):
"""Marlin workspace and act-order sort indices must be recomputed into
the same storage when weights are reloaded (RL weight sync), so device
addresses captured by CUDA graphs remain valid."""
from vllm.model_executor.layers.quantization.utils import marlin_utils
"""Marlin must reuse the workspace storage captured by CUDA graphs."""
_stub_marlin_ops(monkeypatch)
kernel = _make_act_order_marlin_kernel()
generator = torch.Generator().manual_seed(0)
first_g_idx = _random_g_idx(generator)
second_g_idx = _random_g_idx(generator)
kernel = _make_marlin_kernel()
layer = torch.nn.Module()
_load_marlin_checkpoint_format_weights(layer, first_g_idx)
_load_marlin_checkpoint_format_weights(layer)
kernel.process_weights_after_loading(layer)
workspace_ptr = kernel.workspace.data_ptr()
sort_indices_ptr = layer.g_idx_sort_indices.data_ptr()
# Reload: fresh checkpoint-format tensors with a different act-order
_load_marlin_checkpoint_format_weights(layer, second_g_idx)
_load_marlin_checkpoint_format_weights(layer)
kernel.process_weights_after_loading(layer)
assert kernel.workspace.data_ptr() == workspace_ptr
assert torch.all(kernel.workspace == 0)
assert layer.g_idx_sort_indices.data_ptr() == sort_indices_ptr
expected_sort_indices = marlin_utils.marlin_sort_g_idx(second_g_idx)[1]
assert torch.equal(layer.g_idx_sort_indices.data, expected_sort_indices)
# registered as a Parameter so layerwise reload copy-back preserves it
assert isinstance(layer.g_idx_sort_indices, torch.nn.Parameter)
@pytest.mark.parametrize("variant", ["fp8", "mxfp8", "nvfp4"])
@@ -394,7 +361,7 @@ def test_marlin_prepare_layer_preserves_workspace_address(monkeypatch, variant):
monkeypatch.setattr(
ops,
"gptq_marlin_repack",
lambda b_q_weight, perm, size_k, size_n, num_bits, is_a_8bit=False: torch.zeros(
lambda b_q_weight, size_k, size_n, num_bits, is_a_8bit=False: torch.zeros(
size_k // 16, size_n * 2, dtype=torch.int32
),
)
@@ -476,68 +443,6 @@ def test_marlin_make_workspace_new_rejects_incompatible_existing(monkeypatch):
)
def test_marlin_act_order_layerwise_reload_accounting(monkeypatch, dist_init):
"""`g_idx_sort_indices` is generated during weight processing and never
loaded from checkpoints. Registering it as a Parameter must not count it
toward `load_numel_total`: reload restores the construction-time tensor
set before sizing, so act-order layers still process during streaming
instead of deferring (and buffering weights) until finalization."""
from vllm.model_executor.layers.quantization.base_config import (
QuantizeMethodBase,
)
from vllm.model_executor.layers.quantization.utils import marlin_utils
from vllm.model_executor.model_loader.reload.layerwise import get_layerwise_info
_stub_marlin_ops(monkeypatch)
kernel = _make_act_order_marlin_kernel()
class _KernelQuantMethod(QuantizeMethodBase):
def create_weights(self, layer, *args, **kwargs):
raise NotImplementedError
def apply(self, layer, *args, **kwargs):
raise NotImplementedError
def process_weights_after_loading(self, layer):
kernel.process_weights_after_loading(layer)
generator = torch.Generator().manual_seed(0)
layer = torch.nn.Module()
layer.quant_method = _KernelQuantMethod()
_load_marlin_checkpoint_format_weights(layer, _random_g_idx(generator))
# Metadata is recorded at model construction, before any processing
record_metadata_for_reloading(layer)
checkpoint_numel = sum(t.numel() for t in get_layer_tensors(layer).values())
kernel.process_weights_after_loading(layer)
sort_indices = layer.g_idx_sort_indices
initialize_layerwise_reload(layer)
info = get_layerwise_info(layer)
assert info.load_numel_total == checkpoint_numel
# Stream a new checkpoint; the layer must process as soon as its last
# tensor arrives
new_g_idx = _random_g_idx(generator)
checkpoint = {
"qweight": torch.zeros(_MARLIN_SIZE_K // 8, _MARLIN_SIZE_N, dtype=torch.int32),
"scales": torch.ones(
_MARLIN_SIZE_K // _MARLIN_GROUP_SIZE, _MARLIN_SIZE_N, dtype=torch.float16
),
"g_idx": new_g_idx,
}
for name, weight in checkpoint.items():
param = getattr(layer, name)
param.weight_loader(param, weight)
assert not info.can_load()
assert not info.loaded_weights
assert layer.g_idx_sort_indices is sort_indices
expected_sort_indices = marlin_utils.marlin_sort_g_idx(new_g_idx)[1]
assert torch.equal(layer.g_idx_sort_indices.data, expected_sort_indices)
def test_model_cleanup(dist_init, default_vllm_config):
layer = QKVParallelLinear(2, 3, 4)
assert layer.weight.weight_loader.__self__ is layer