[Refactor] Stop passing models' layers the placement they already read (#41815)

This commit is contained in:
Cheng Wan
2026-09-30 13:28:17 -07:00
committed by GitHub
parent c7fa37a5ca
commit 00659ac099
6 changed files with 11 additions and 32 deletions
+2 -7
View File
@@ -779,8 +779,6 @@ class MqaAttentionBase(nn.Module):
quant_config: Optional[QuantizationConfig],
prefix: str,
*,
attn_tp_rank: Optional[int] = None,
attn_tp_size: Optional[int] = None,
compress_ratio: Optional[int] = None,
fuse_wqa_wkv: Optional[bool] = None,
wo_a_fp8: Optional[bool] = None,
@@ -791,11 +789,8 @@ class MqaAttentionBase(nn.Module):
super().__init__()
self.is_dsv41 = getattr(config, "model_type", None) == "deepseek_v41"
self.dsa_enable_prefill_cp = is_dsa_enable_prefill_cp()
if attn_tp_rank is None or attn_tp_size is None:
attn_tp_rank = get_parallel().attn_tp_rank
attn_tp_size = get_parallel().attn_tp_size
self.attn_tp_rank: int = attn_tp_rank
self.attn_tp_size: int = attn_tp_size
self.attn_tp_rank: int = get_parallel().attn_tp_rank
self.attn_tp_size: int = get_parallel().attn_tp_size
self.layer_id = layer_id
self.dim = config.hidden_size
@@ -109,8 +109,6 @@ class DSparkAttention(MqaAttentionBase):
layer_id,
quant_config,
prefix,
attn_tp_rank=get_parallel().attn_tp_rank,
attn_tp_size=get_parallel().attn_tp_size,
compress_ratio=0,
fuse_wqa_wkv=False,
wo_a_fp8=False,
+2 -4
View File
@@ -1,7 +1,6 @@
import logging
from array import array
from contextlib import nullcontext
from functools import partial
from typing import Iterable, List, Optional, Tuple, Union
import torch
@@ -393,7 +392,6 @@ class Glm5NextLinearAttention(nn.Module):
super().__init__()
head_shard_size = get_parallel().attn_tp_size
head_shard_rank = get_parallel().attn_tp_rank
_head_shard_rank_getter = partial(getattr, get_parallel(), "attn_tp_rank")
self.hidden_size = hidden_size
self.config = config
@@ -532,7 +530,7 @@ class Glm5NextLinearAttention(nn.Module):
set_weight_attrs(
self.dt_bias,
{"weight_loader": sharded_weight_loader(0, _head_shard_rank_getter)},
{"weight_loader": sharded_weight_loader(0)},
)
self.qkv_conv1d = MergedColumnParallelLinear(
@@ -553,7 +551,7 @@ class Glm5NextLinearAttention(nn.Module):
)
set_weight_attrs(
self.A_log,
{"weight_loader": sharded_weight_loader(2, _head_shard_rank_getter)},
{"weight_loader": sharded_weight_loader(2)},
)
self.o_norm = FusedRMSNormGated(
+7 -12
View File
@@ -331,21 +331,20 @@ class Grok1Attention(nn.Module):
self.config = config
self.layer_id = layer_id
self.hidden_size = hidden_size
attn_tp_rank = get_parallel().tp_rank
attn_tp_size = get_parallel().tp_size
tp_size = get_parallel().tp_size
self.total_num_heads = num_heads
assert self.total_num_heads % attn_tp_size == 0
self.num_heads = self.total_num_heads // attn_tp_size
assert self.total_num_heads % tp_size == 0
self.num_heads = self.total_num_heads // tp_size
self.total_num_kv_heads = num_kv_heads
if self.total_num_kv_heads >= attn_tp_size:
if self.total_num_kv_heads >= tp_size:
# Number of KV heads is greater than TP size, so we partition
# the KV heads across multiple tensor parallel GPUs.
assert self.total_num_kv_heads % attn_tp_size == 0
assert self.total_num_kv_heads % tp_size == 0
else:
# Number of KV heads is less than TP size, so we replicate
# the KV heads across multiple tensor parallel GPUs.
assert attn_tp_size % self.total_num_kv_heads == 0
self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size)
assert tp_size % self.total_num_kv_heads == 0
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
self.head_dim = getattr(config, "head_dim", 128)
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
@@ -362,8 +361,6 @@ class Grok1Attention(nn.Module):
self.total_num_kv_heads,
bias=False,
quant_config=quant_config,
tp_rank=attn_tp_rank,
tp_size=attn_tp_size,
load_presharded_attn=self.load_presharded_attn,
prefix=add_prefix("qkv_proj", prefix),
)
@@ -373,8 +370,6 @@ class Grok1Attention(nn.Module):
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
tp_rank=attn_tp_rank,
tp_size=attn_tp_size,
use_presharded_weights=self.load_presharded_attn,
prefix=add_prefix("o_proj", prefix),
)
-2
View File
@@ -294,8 +294,6 @@ class IQuestQ1Attention(nn.Module):
quant_config=quant_config,
tp_rank=attn_tp_rank,
tp_size=attn_tp_size,
kv_tp_rank=attn_tp_rank,
kv_tp_size=attn_tp_size,
prefix=add_prefix("qkv_proj", prefix),
)
self.o_proj = RowParallelLinear(
-5
View File
@@ -829,7 +829,6 @@ class ZayaAttention(nn.Module):
# divisible by tp_size; the KV-replicated GQA-TP variant (tp_size >
# num_k_heads) is intentionally rejected with a clear error message
# because both per-K-head paths assume each rank holds whole K heads.
self.tp_rank = get_parallel().tp_rank
self.tp_size = get_parallel().tp_size
# The head split, the ``o_proj`` RowParallel all-reduce, and the
# RadixAttention KV cache are all organized on the *global* TP group,
@@ -873,8 +872,6 @@ class ZayaAttention(nn.Module):
layer_id=layer_id,
quant_config=quant_config,
prefix=add_prefix("qkv", prefix),
tp_rank=self.tp_rank,
tp_size=self.tp_size,
)
# RowParallel o_proj: per-rank input is the rank's q heads, full
@@ -887,8 +884,6 @@ class ZayaAttention(nn.Module):
reduce_results=True,
quant_config=quant_config,
prefix=add_prefix("o_proj", prefix),
tp_rank=self.tp_rank,
tp_size=self.tp_size,
)
rope_theta = float(getattr(config, "rope_theta", 1_000_000.0))