[Bugfix] Preserve Marlin runtime tensor storage across weight reload (#48438)

Signed-off-by: Ryan Clark <ryanclark2k@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Ryan Clark
2026-07-30 14:50:03 -07:00
committed by GitHub
co-authored by Claude Fable 5
parent e9096fcb12
commit 68fb303a41
6 changed files with 330 additions and 17 deletions
@@ -167,6 +167,283 @@ def test_materialize_layer_preserves_non_meta_tensors():
assert torch.equal(layer.bias.data, bias_values)
_MARLIN_SIZE_K, _MARLIN_SIZE_N, _MARLIN_GROUP_SIZE = 128, 64, 64
def _stub_marlin_ops(monkeypatch):
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.utils import marlin_utils
monkeypatch.setattr(marlin_utils, "num_compute_units", lambda _: 4)
monkeypatch.setattr(
ops,
"gptq_marlin_repack",
lambda w, perm, 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():
from vllm.model_executor.kernels.linear.mixed_precision.marlin import (
MarlinLinearKernel,
)
from vllm.model_executor.kernels.linear.mixed_precision.MPLinearKernel import (
MPLinearLayerConfig,
)
from vllm.scalar_type import scalar_types
kernel = object.__new__(MarlinLinearKernel)
kernel.config = MPLinearLayerConfig(
full_weight_shape=(_MARLIN_SIZE_K, _MARLIN_SIZE_N),
partition_weight_shape=(_MARLIN_SIZE_K, _MARLIN_SIZE_N),
weight_type=scalar_types.uint4b8,
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):
from vllm.model_executor.parameter import (
GroupQuantScaleParameter,
PackedvLLMParameter,
RowvLLMParameter,
)
layer.qweight = PackedvLLMParameter(
data=torch.zeros(_MARLIN_SIZE_K // 8, _MARLIN_SIZE_N, dtype=torch.int32),
input_dim=0,
output_dim=1,
packed_dim=0,
packed_factor=8,
weight_loader=default_weight_loader,
)
layer.scales = GroupQuantScaleParameter(
data=torch.ones(
_MARLIN_SIZE_K // _MARLIN_GROUP_SIZE, _MARLIN_SIZE_N, dtype=torch.float16
),
input_dim=0,
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
_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)
layer = torch.nn.Module()
_load_marlin_checkpoint_format_weights(layer, first_g_idx)
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)
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"])
def test_marlin_prepare_layer_preserves_workspace_address(monkeypatch, variant):
"""The Marlin fallback prepare_* functions rerun on weight reload and must
reuse the workspace storage whose address captured CUDA graphs hold."""
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.utils import (
marlin_utils,
marlin_utils_fp4,
marlin_utils_fp8,
)
size_k, size_n = 128, 64
monkeypatch.setattr(marlin_utils, "num_compute_units", lambda _: 4)
monkeypatch.setattr(
ops,
"gptq_marlin_repack",
lambda b_q_weight, perm, size_k, size_n, num_bits, is_a_8bit=False: torch.zeros(
size_k // 16, size_n * 2, dtype=torch.int32
),
)
layer = torch.nn.Module()
layer.output_size_per_partition = size_n
layer.input_size_per_partition = size_k
layer.orig_dtype = torch.float16
layer.params_dtype = torch.float16
if variant == "fp8":
prepare = marlin_utils_fp8.prepare_fp8_layer_for_marlin
def load_checkpoint_format_weights():
layer.weight = torch.nn.Parameter(
torch.zeros(size_k, size_n, dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(
torch.ones(1, dtype=torch.float32), requires_grad=False
)
elif variant == "mxfp8":
prepare = marlin_utils_fp8.prepare_mxfp8_layer_for_marlin
def load_checkpoint_format_weights():
layer.weight = torch.nn.Parameter(
torch.zeros(size_n, size_k, dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(
torch.full((size_n, size_k // 32), 127, dtype=torch.uint8),
requires_grad=False,
)
else:
prepare = marlin_utils_fp4.prepare_fp4_layer_for_marlin
def load_checkpoint_format_weights():
layer.weight = torch.nn.Parameter(
torch.zeros(size_n, size_k // 2, dtype=torch.uint8),
requires_grad=False,
)
layer.weight_scale = torch.nn.Parameter(
torch.ones(size_n, size_k // 16, dtype=torch.float8_e4m3fn),
requires_grad=False,
)
layer.weight_global_scale = torch.nn.Parameter(
torch.ones(1, dtype=torch.float32), requires_grad=False
)
load_checkpoint_format_weights()
prepare(layer)
workspace_ptr = layer.workspace.data_ptr()
# Reload: fresh checkpoint-format tensors, prepare runs again
load_checkpoint_format_weights()
prepare(layer)
assert layer.workspace.data_ptr() == workspace_ptr
assert torch.all(layer.workspace == 0)
def test_marlin_make_workspace_new_rejects_incompatible_existing(monkeypatch):
"""An incompatible existing workspace means the address captured by CUDA
graphs is already unusable; allocating a replacement would hide that."""
from vllm.model_executor.layers.quantization.utils import marlin_utils
monkeypatch.setattr(marlin_utils, "num_compute_units", lambda _: 4)
device = torch.device("cpu")
workspace = marlin_utils.marlin_make_workspace_new(device)
reused = marlin_utils.marlin_make_workspace_new(device, existing=workspace)
assert reused is workspace
with pytest.raises(ValueError, match="incompatible"):
marlin_utils.marlin_make_workspace_new(device, 4, existing=workspace)
with pytest.raises(ValueError, match="incompatible"):
marlin_utils.marlin_make_workspace_new(
device, existing=workspace.to(torch.int64)
)
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
@@ -25,6 +25,7 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
unpack_cols,
)
from vllm.model_executor.parameter import BasevLLMParameter, permute_param_layout_
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.scalar_type import scalar_types
@@ -111,8 +112,10 @@ class MarlinLinearKernel(MPLinearKernel):
else:
padded_n, padded_k = marlin_padded_nk(size_n, size_k, c.group_size)
# Allocate marlin workspace.
self.workspace = marlin_make_workspace_new(device)
# Allocate marlin workspace, reusing existing storage on reload.
self.workspace = marlin_make_workspace_new(
device, existing=getattr(self, "workspace", None)
)
# Default names since marlin requires empty parameters for these,
# TODO: remove this requirement from marlin (allow optional tensors)
@@ -174,7 +177,9 @@ class MarlinLinearKernel(MPLinearKernel):
getattr(layer, self.w_gidx_name)
)
self._transform_param(layer, self.w_gidx_name, lambda _: g_idx)
layer.g_idx_sort_indices = g_idx_sort_indices
replace_parameter(
layer, "g_idx_sort_indices", g_idx_sort_indices, prefer_copy=True
)
else:
setattr(layer, self.w_gidx_name, marlin_make_empty_g_idx(device))
layer.g_idx_sort_indices = marlin_make_empty_g_idx(device)
@@ -504,7 +504,9 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
and self.wna16_backend != WNA16MoEBackend.EMULATION
):
layer.workspace = marlin_make_workspace_new(
layer.w13_weight_g_idx.device, 4
layer.w13_weight_g_idx.device,
4,
existing=getattr(layer, "workspace", None),
)
# Alias packed weights to w13_weight/w2_weight for the modular kernel interface
@@ -397,14 +397,31 @@ def marlin_moe_intermediate_size(w1_packed: torch.Tensor, w2_packed: torch.Tenso
def marlin_make_workspace_new(
device: torch.device, max_blocks_per_sm: int = 1
device: torch.device,
max_blocks_per_sm: int = 1,
existing: torch.Tensor | None = None,
) -> torch.Tensor:
# In the new marlin kernel, we use the num of threadblocks as workspace
# size. The num of threadblocks is sms_count * max_blocks_per_sm.
sms = num_compute_units(device.index)
return torch.zeros(
sms * max_blocks_per_sm, dtype=torch.int, device=device, requires_grad=False
)
size = sms * max_blocks_per_sm
# On weight reload, reuse the existing storage so the workspace address
# captured by CUDA graphs stays valid.
if existing is not None:
if (
existing.device != device
or existing.dtype != torch.int
or existing.numel() != size
):
raise ValueError(
f"Existing Marlin workspace is incompatible "
f"(device={existing.device}, dtype={existing.dtype}, "
f"numel={existing.numel()}; expected device={device}, "
f"dtype={torch.int}, numel={size}). Reload must reuse the "
f"workspace storage captured by CUDA graphs."
)
return existing.zero_()
return torch.zeros(size, dtype=torch.int, device=device, requires_grad=False)
def marlin_is_k_full(act_order: bool, is_row_parallel: bool) -> bool:
@@ -240,7 +240,9 @@ def prepare_fp4_layer_for_marlin(
device = layer.weight.device
# WORKSPACE
layer.workspace = marlin_make_workspace_new(device)
layer.workspace = marlin_make_workspace_new(
device, existing=getattr(layer, "workspace", None)
)
# WEIGHT
# Repack weights to marlin format
@@ -396,7 +398,9 @@ def prepare_nvfp4_moe_layer_for_marlin(
is_a_8bit = input_dtype is not None and input_dtype.itemsize == 1
# WORKSPACE
layer.workspace = marlin_make_workspace_new(device, 4)
layer.workspace = marlin_make_workspace_new(
device, 4, existing=getattr(layer, "workspace", None)
)
perm = torch.empty(0, dtype=torch.int, device=device)
# WEIGHT
@@ -485,7 +489,9 @@ def prepare_moe_fp4_layer_for_marlin(
# WORKSPACE
device = layer.w13_weight.device
param_dtype = layer.params_dtype
layer.workspace = marlin_make_workspace_new(device, 4)
layer.workspace = marlin_make_workspace_new(
device, 4, existing=getattr(layer, "workspace", None)
)
perm = torch.empty(0, dtype=torch.int, device=device)
is_a_8bit = input_dtype is not None and input_dtype.itemsize == 1
@@ -132,7 +132,9 @@ def prepare_fp8_layer_for_marlin(
device = layer.weight.device
# WORKSPACE
layer.workspace = marlin_make_workspace_new(device)
layer.workspace = marlin_make_workspace_new(
device, existing=getattr(layer, "workspace", None)
)
# WEIGHT
# Repack weights to marlin format
@@ -278,9 +280,9 @@ def prepare_fp8_moe_layer_for_marlin(
# WORKSPACE
device = layer.w13_weight.device
# NOTE(rob): we do not need to register the workspace as a param
# because it is not used as part of the weight reloading process.
layer.workspace = marlin_make_workspace_new(device, 4)
layer.workspace = marlin_make_workspace_new(
device, 4, existing=getattr(layer, "workspace", None)
)
perm = torch.empty(0, dtype=torch.int, device=device)
# WEIGHT
@@ -463,7 +465,9 @@ def prepare_mxfp8_layer_for_marlin(layer: torch.nn.Module) -> None:
device = layer.weight.device
# WORKSPACE
layer.workspace = marlin_make_workspace_new(device)
layer.workspace = marlin_make_workspace_new(
device, existing=getattr(layer, "workspace", None)
)
# WEIGHT - repack FP8 weights to Marlin format
perm = torch.empty(0, dtype=torch.int, device=device)
@@ -549,7 +553,9 @@ def prepare_mxfp8_moe_layer_for_marlin(
param_dtype = torch.get_default_dtype()
perm = torch.empty(0, dtype=torch.int, device=device)
layer.workspace = marlin_make_workspace_new(device, 4)
layer.workspace = marlin_make_workspace_new(
device, 4, existing=getattr(layer, "workspace", None)
)
def repack_weight(weight: torch.Tensor, name: str) -> torch.Tensor:
if "w13" in name: