[Bugfix] Load stacked expert weights for non-gated MoE (#59031)

Signed-off-by: Roi Koren <roik@nvidia.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
roikoren755
2026-10-05 15:26:55 +03:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 4ff028d77e
commit 51eeb0c58f
6 changed files with 88 additions and 18 deletions
@@ -9,6 +9,8 @@ have the original unpadded size. These tests verify that weight loading
correctly handles this mismatch.
"""
from types import SimpleNamespace
import pytest
import torch
@@ -564,9 +566,11 @@ class TestLoadWeightsExpertMapping:
transposed=False,
quant_method="tensor",
layer_name="model.layers.0.mlp.experts",
is_gated=True,
):
experts = torch.nn.Module()
experts.layer_name = layer_name
experts.moe_config = SimpleNamespace(is_act_and_mul=is_gated)
experts.get_expert_mapping = lambda **_: mapping
experts.is_fused_checkpoint_transposed = transposed
experts._orient_fused_weight = RoutedExperts._orient_fused_weight
@@ -593,7 +597,7 @@ class TestLoadWeightsExpertMapping:
(("gate_proj", "down_proj", "up_proj"), ""),
(("w1", "w2", "w3"), ""),
(("up_proj", "down_proj", "up_proj"), ""),
(("up_proj", "down_proj", ""), ""),
(("up_proj", "down_proj", None), ""),
(("gate_proj", "down_proj", "up_proj"), "base_layer."),
],
)
@@ -609,7 +613,10 @@ class TestLoadWeightsExpertMapping:
self, projs, lora_prefix, suffix, shape, dtype
):
mapping = self._mapping(
projs, num_redundant_experts=4, lora_base_layer_prefix=lora_prefix
projs,
num_redundant_experts=4,
lora_base_layer_prefix=lora_prefix,
is_gated=projs[2] is not None,
)
for logical_id, physical_ids in [(1, (1, 17)), (10, (10,))]:
for proj in dict.fromkeys(projs):
@@ -651,6 +658,47 @@ class TestLoadWeightsExpertMapping:
for call, expected in zip(calls, weight.chunk(2) * 2):
torch.testing.assert_close(call[3], expected, rtol=0, atol=0)
def test_non_gated_mapping_uses_stacked_up_proj_without_warning(self, caplog_vllm):
mapping = self._mapping(("up_proj", "down_proj", None), is_gated=False)
assert "Unexpected gate/up projection names" not in caplog_vllm.text
assert mapping[:2] == [
("experts.w13_weight", "experts.up_proj", 0, "w1"),
("experts.w2_weight", "experts.down_proj", 0, "w2"),
]
assert mapping[2:] == [
(param, f"experts.{expert}.{proj}.", expert, shard)
for expert in range(16)
for param, proj, shard in (
("experts.w13_", "up_proj", "w1"),
("experts.w2_", "down_proj", "w2"),
)
]
def test_gated_mapping_with_unknown_names_warns_and_skips_fused(self, caplog_vllm):
mapping = self._mapping(("fc1", "fc2", "fc3"))
assert "Unexpected gate/up projection names: fc1, fc3" in caplog_vllm.text
assert [name for _, name, _, _ in mapping[:3]] == [
"experts.0.fc1.",
"experts.0.fc2.",
"experts.0.fc3.",
]
def test_stacked_non_gated_tensors_load_unsplit(self):
mapping = self._mapping(("up_proj", "down_proj", None), is_gated=False)
up_proj = torch.arange(16 * 4 * 3).reshape(16, 4, 3)
down_proj = torch.arange(16 * 3 * 4).reshape(16, 3, 4)
calls = self._load(
mapping,
[("up_proj", up_proj), ("down_proj", down_proj)],
is_gated=False,
)
assert [call[:3] for call in calls] == [
*(("w13_weight", "w1", expert) for expert in range(16)),
*(("w2_weight", "w2", expert) for expert in range(16)),
]
for call, expected in zip(calls, [*up_proj, *down_proj]):
torch.testing.assert_close(call[3], expected, rtol=0, atol=0)
@pytest.mark.parametrize(
"suffix,quant_method,transposed",
[
+2 -4
View File
@@ -409,13 +409,11 @@ def process_packed_modules_mapping(
"experts.w3",
]
elif (not model.is_3d_moe_weight) or force_2d_moe:
# Filter out malformed entries: non-gated MoE has empty
# ckpt_up_proj_name which results in weight_name containing ".."
# (e.g., "experts.0.." instead of "experts.0.layer_name.")
# Non-gated MoE yields two entries per expert (no w3); the
# manager pads them to triplets before packing.
packed_modules_mapping["experts"] = [
weight_name.rstrip(".")
for _, weight_name, _, _ in get_moe_expert_mapping(model)
if ".." not in weight_name
]
return packed_modules_mapping
@@ -118,7 +118,7 @@ def FusedMoEFactory(
has_bias: bool = False,
is_sequence_parallel: bool = False,
reduce_results: bool = True,
ckpt_names: tuple[str, str, str] = ("gate_proj", "down_proj", "up_proj"),
ckpt_names: tuple[str, str, str | None] = ("gate_proj", "down_proj", "up_proj"),
is_fused_checkpoint_transposed: bool = False,
n_shared_experts: int | None = None,
fuse_shared_experts: bool = False,
@@ -442,7 +442,7 @@ def fused_moe_make_expert_params_mapping(
model: torch.nn.Module,
ckpt_gate_proj_name: str,
ckpt_down_proj_name: str,
ckpt_up_proj_name: str,
ckpt_up_proj_name: str | None,
num_experts: int,
num_redundant_experts: int = 0,
routed_experts_prefix: str = "routed_experts",
@@ -87,7 +87,7 @@ class RoutedExperts(PluggableLayer):
expert_map_manager: ExpertMapManager,
ckpt_gate_proj_name: str = "gate_proj",
ckpt_down_proj_name: str = "down_proj",
ckpt_up_proj_name: str = "up_proj",
ckpt_up_proj_name: str | None = "up_proj",
is_fused_checkpoint_transposed: bool = False,
#
# Extra params that are needed by quant_methods, pass along for now
@@ -114,6 +114,13 @@ class RoutedExperts(PluggableLayer):
self.ckpt_gate_proj_name = ckpt_gate_proj_name
self.ckpt_down_proj_name = ckpt_down_proj_name
self.ckpt_up_proj_name = ckpt_up_proj_name
if moe_config.is_act_and_mul == (ckpt_up_proj_name is None):
raise ValueError(
f"{layer_name}: ckpt_up_proj_name={ckpt_up_proj_name!r} is "
f"inconsistent with activation {moe_config.activation.value!r}. "
"Gated MoE needs an up projection name; non-gated MoE must "
"pass None."
)
self.is_fused_checkpoint_transposed = is_fused_checkpoint_transposed
self.expert_map_manager = expert_map_manager
self.hidden_size = moe_config.hidden_dim
@@ -971,7 +978,7 @@ class RoutedExperts(PluggableLayer):
loaded_weight,
self.is_fused_checkpoint_transposed and uses_weight_layout,
)
if shard_id in {"w1", "w3"}:
if shard_id in {"w1", "w3"} and self.moe_config.is_act_and_mul:
# Repurpose expert_id for deconcatenating w1 and w3
experts_shard = fused_weight.chunk(2, dim=1)[expert_id]
else:
@@ -1029,6 +1036,7 @@ class RoutedExperts(PluggableLayer):
routed_experts_prefix="",
lora_base_layer_prefix=self.lora_base_layer_prefix,
include_fused=include_fused,
is_gated=moe_config.is_act_and_mul,
)
@staticmethod
@@ -1036,7 +1044,7 @@ class RoutedExperts(PluggableLayer):
model: torch.nn.Module,
ckpt_gate_proj_name: str,
ckpt_down_proj_name: str,
ckpt_up_proj_name: str,
ckpt_up_proj_name: str | None,
num_experts: int,
num_redundant_experts: int = 0,
routed_experts_prefix: str = "routed_experts",
@@ -1067,13 +1075,14 @@ class RoutedExperts(PluggableLayer):
def build_expert_params_mapping(
ckpt_gate_proj_name: str,
ckpt_down_proj_name: str,
ckpt_up_proj_name: str,
ckpt_up_proj_name: str | None,
num_experts: int,
num_redundant_experts: int = 0,
routed_experts_prefix: str = "routed_experts",
lora_base_layer_prefix: str = "",
lora_base_layer_prefix_on_param_name: str = "",
include_fused: bool = False,
is_gated: bool = True,
) -> list[tuple[str, str, int, str]]:
"""Create expert parameter mapping for weight loading with redundant experts.
@@ -1095,6 +1104,8 @@ class RoutedExperts(PluggableLayer):
``make_expert_params_mapping`` indexes the model-wide
``params_dict`` (prefix included).
include_fused: Prepend the fused pre-fused-checkpoint entries
is_gated: Whether w13 holds gate and up (gated activation) or a
single up projection. Selects the fused checkpoint layout.
routed_experts_prefix: Prefix of the routed experts submodule
Returns:
@@ -1129,7 +1140,13 @@ class RoutedExperts(PluggableLayer):
fused_mapping = []
if include_fused:
gate_up = None
if ckpt_gate_proj_name == "gate_proj" and ckpt_up_proj_name == "up_proj":
w13_shards: tuple[str, ...] = ("w1", "w3")
if not is_gated:
# Non-gated: the stacked checkpoint tensor is the up projection
# itself, nothing to split into gate and up.
gate_up = ckpt_gate_proj_name
w13_shards = ("w1",)
elif ckpt_gate_proj_name == "gate_proj" and ckpt_up_proj_name == "up_proj":
gate_up = "gate_up_proj"
elif ckpt_gate_proj_name == "w1" and ckpt_up_proj_name == "w3":
gate_up = "w13"
@@ -1143,10 +1160,16 @@ class RoutedExperts(PluggableLayer):
if gate_up is not None:
fused_mapping = [
# (param_name, weight_name, expert_id, shard_id)
(f"{w13}weight", f"experts.{gate_up}", 0, "w1"),
(f"{w13}weight", f"experts.{gate_up}", 1, "w3"),
# expert_id doubles as the chunk index when splitting w13.
*(
(f"{w13}weight", f"experts.{gate_up}", chunk, shard_id)
for chunk, shard_id in enumerate(w13_shards)
),
(f"{w2}weight", f"experts.{ckpt_down_proj_name}", 0, "w2"),
]
if gate_up is not None and is_gated:
# Per-expert gate_up_proj tensors; the non-gated equivalent is
# already covered by the per-expert mapping below.
fused_mapping.extend(
(
w13,
@@ -1175,6 +1198,7 @@ class RoutedExperts(PluggableLayer):
("w2", ckpt_down_proj_name),
("w3", ckpt_up_proj_name),
]
if weight_name is not None
]
return fused_mapping + per_expert_mapping
+2 -2
View File
@@ -235,7 +235,7 @@ class NemotronHMoE(nn.Module):
intermediate_size=config.moe_intermediate_size,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
ckpt_names=("up_proj", "down_proj", ""),
ckpt_names=("up_proj", "down_proj", None),
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
@@ -707,7 +707,7 @@ class NemotronHModel(nn.Module, EagleModelMixin):
self,
ckpt_gate_proj_name="up_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="",
ckpt_up_proj_name=None,
num_experts=self._get_max_n_routed_experts(),
num_redundant_experts=getattr(self, "num_redundant_experts", 0),
)
+1 -1
View File
@@ -444,7 +444,7 @@ class NemotronHMTP(nn.Module, SupportsPP, SupportsQuant):
self,
ckpt_gate_proj_name="up_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="", # Empty - non-gated MoE
ckpt_up_proj_name=None, # non-gated MoE
num_experts=num_experts,
num_redundant_experts=self.num_redundant_experts,
)