[Bugfix] Re-sync parameter tp_rank after process_weights_after_loading (fix replicated / disable_tp weight reload) (#48025)

Signed-off-by: Alex Xu <alexxu@roblox.com>
Co-authored-by: YQ-Wang <yiqingwang@roblox.com>
Co-authored-by: alexhxu <alex.xu1015@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
alexxu-roblox
2026-07-18 08:40:06 +00:00
committed by GitHub
co-authored by YQ-Wang alexhxu Cursor
parent c71a583aa9
commit d96aee0951
3 changed files with 23 additions and 7 deletions
+11 -7
View File
@@ -281,6 +281,15 @@ class LinearBase(PluggableLayer):
self.tp_size = get_tensor_model_parallel_world_size() if not disable_tp else 1
def update_param_tp_status(self):
# Single source of truth for a parameter's TP state. BasevLLMParameter
# stamps self.tp_rank with the *global* rank in __init__; this reconciles
# every child parameter to the *layer's* tp_rank/tp_size (which correctly
# accounts for disable_tp -> replicated weights with tp_rank == 0).
#
# Must be re-run whenever parameters are (re-)created after construction,
# e.g. after quant_method.process_weights_after_loading() swaps in fresh
# Parameters. Otherwise a later load_weights()/weight-refit would narrow a
# replicated weight at global_rank * shard_size and overflow.
for param in self.parameters():
if isinstance(param, BasevLLMParameter):
param.tp_rank = self.tp_rank
@@ -910,7 +919,6 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
shard_id=loaded_shard_id,
shard_offset=shard_offset,
shard_size=shard_size,
tp_rank=self.tp_rank,
)
def load_weights(
@@ -1110,12 +1118,10 @@ class QKVParallelLinear(ColumnParallelLinear):
# to ensure that any subsequent reduction (like .max())
# works correctly while preserving the parameter shape.
for idx in range(param.data.shape[0]):
param.load_qkv_weight(
loaded_weight=loaded_weight, shard_id=idx, tp_rank=self.tp_rank
)
param.load_qkv_weight(loaded_weight=loaded_weight, shard_id=idx)
return
elif type(param) in (RowvLLMParameter, BasevLLMParameter):
param.load_qkv_weight(loaded_weight=loaded_weight, tp_rank=self.tp_rank)
param.load_qkv_weight(loaded_weight=loaded_weight)
return
# TODO: @dsikka - move to parameter.py
self._load_fused_module_from_checkpoint(param, loaded_weight)
@@ -1139,7 +1145,6 @@ class QKVParallelLinear(ColumnParallelLinear):
shard_id=loaded_shard_id,
shard_offset=shard_offset,
shard_size=shard_size,
tp_rank=self.tp_rank,
)
def weight_loader(
@@ -1493,7 +1498,6 @@ class MinimaxM3QKVParallelLinearWithIndexer(QKVParallelLinear):
shard_id=loaded_shard_id,
shard_offset=shard_offset,
shard_size=shard_size,
tp_rank=self.tp_rank,
)
def weight_loader(
@@ -364,6 +364,11 @@ def _layerwise_process(layer: torch.nn.Module, info: LayerReloadingInfo):
quant_method = getattr(layer, "quant_method", None)
if isinstance(quant_method, QuantizeMethodBase):
quant_method.process_weights_after_loading(layer)
# Re-reconcile parameter TP state: process_weights_after_loading may
# have re-created Parameters (stamped with the global rank), which would
# otherwise break replicated (disable_tp) weights on a subsequent reload.
if hasattr(layer, "update_param_tp_status"):
layer.update_param_tp_status()
# Copy processed values into original tensor storage (preserves cudagraph refs)
# this code is a no-op if not reloading (because kernel tensors is empty)
@@ -111,6 +111,13 @@ def process_weights_after_loading(
# parameters onto device for processing and back off after.
with device_loading_context(module, target_device):
quant_method.process_weights_after_loading(module)
# process_weights_after_loading may swap in freshly-created
# Parameters (e.g. FP8 requantization), which are stamped with the
# global rank in BasevLLMParameter.__init__. Re-reconcile their TP
# state to the layer so a later weight reload / RL weight-refit
# narrows replicated (disable_tp) weights at the correct offset.
if hasattr(module, "update_param_tp_status"):
module.update_param_tp_status()
# Repacking transients above can leave large amounts of memory in
# the caching allocator, which starves the OS on UMA devices.
release_device_memory_under_pressure(target_device)