[Transformers backend] Find attention with a fuser and attach vLLM's layer to it (#54941)

Signed-off-by: Thomas Ortner <boh@zurich.ibm.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Thomas Ortner
2026-09-04 16:26:01 +00:00
committed by GitHub
co-authored by Harry Mellor
parent 99a1ab8e4c
commit 8ad2076c4a
14 changed files with 521 additions and 150 deletions
+78 -34
View File
@@ -57,7 +57,11 @@ from vllm.model_executor.models.interfaces import (
)
from vllm.model_executor.models.interfaces_base import VllmModel
from vllm.model_executor.models.transformers.fuser import BaseFuser, Fusers
from vllm.model_executor.models.transformers.fusers import MLAFuser
from vllm.model_executor.models.transformers.fusers import AttentionFuser, MLAFuser
from vllm.model_executor.models.transformers.fusers.attention import (
VLLM_ATTN_IMPL,
VLLM_MLA_ATTN_IMPL,
)
from vllm.model_executor.models.transformers.utils import (
attrsetter,
can_enable_torch_compile,
@@ -73,7 +77,6 @@ from vllm.model_executor.models.utils import (
AutoWeightsLoader,
PPMissingLayer,
WeightsMapper,
extract_layer_index,
make_empty_intermediate_tensors_factory,
maybe_prefix,
)
@@ -86,6 +89,11 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
VLLM_ATTN_ATTR = "attn"
"""Attribute in which to store the vLLM `Attention` instance that serves a Transformers
attention module. `"attn"` is chosen deliberately to match the mappings generated by
`maybe_remap_kv_scale_name`."""
class PreTrainedModelClasses(NamedTuple):
decoder: type["PreTrainedModel"]
@@ -129,9 +137,13 @@ class Base(
self.packed_modules_mapping: dict[str, list[str]] = {}
"""Fused module -> constituent projections, populated by `recursive_replace`
for the quantization machinery and loaders (e.g. bitsandbytes)."""
self.fusers: dict[str, BaseFuser] = {}
"""Module qualname -> the fuser applied to it, populated
self.fusers: dict[str, list[BaseFuser]] = {}
"""Module qualname -> the fusers applied to it, populated
by `recursive_replace` for `create_attention_instances`."""
self.attention_fusers: dict[int, tuple[str, AttentionFuser]] = {}
"""`layer_idx` -> the qualname and fuser of the module computing that
layer's attention, populated by `recursive_replace` for
`create_attention_instances`."""
# Attrs for Eagle3 (see self.set_aux_hidden_state_layers)
self._target_class: type[nn.Module] = nn.Module
@@ -189,7 +201,7 @@ class Base(
- Sets the dtype to the default torch dtype set by vLLM because Transformers
uses the config dtype when creating the model
"""
self.text_config._attn_implementation = "vllm"
self.text_config._attn_implementation = VLLM_ATTN_IMPL
self.config.dtype = torch.get_default_dtype()
@contextmanager
@@ -475,9 +487,22 @@ class Base(
vocab_embeddings = self._vocab_embeddings()
def register_fusion(fuser: BaseFuser, prefix: str):
def register_fusion(fuser: BaseFuser, prefix: str, module: nn.Module):
"""Register a fused layer's mappings just before it is built."""
self.fusers[prefix] = fuser
self.fusers.setdefault(prefix, []).append(fuser)
if (
isinstance(fuser, AttentionFuser)
and (index := fuser.layer_index(module)) is not None
):
if index in self.attention_fusers:
other, _ = self.attention_fusers[index]
raise ValueError(
f"Layer {index} dispatches attention from more than one "
f"module ({other} and {prefix}), which one vLLM attention "
"layer cannot serve."
)
self.attention_fusers[index] = (prefix, fuser)
orig_to_new_stacked = fuser.orig_to_new_stacked(prefix)
self.hf_to_vllm_mapper.orig_to_new_stacked.update(orig_to_new_stacked)
@@ -528,10 +553,11 @@ class Base(
new_module = replace_embedding_class(
child_module, self.quant_config, prefix=qual_name
)
elif (fuser := fusers[child_module]) is not None:
register_fusion(fuser, qual_name)
new_module = fuser.fuse(child_module, qual_name, self.vllm_config)
logger.info_once(fuser.info(child_name))
elif child_module_fusers := fusers[child_module]:
for fuser in child_module_fusers:
register_fusion(fuser, qual_name, child_module)
new_module = fuser.fuse(new_module, qual_name, self.vllm_config)
logger.info_once(fuser.info(child_name))
_recursive_replace(new_module, prefix=qual_name)
elif not isinstance(child_module, MoERunner):
# MoERunner can contain aliases of shared experts and gates,
@@ -548,20 +574,14 @@ class Base(
"""
Create `Attention` instances to inform KV cache allocation.
"""
mla_fusers = {}
attention_instances = {}
text_config = self.text_config
attn_cls = self._get_attn_cls()
# kv_lora_rank indicates that this is an MLA model
if getattr(text_config, "kv_lora_rank", None) is not None:
mla_fusers = {
extract_layer_index(prefix): (prefix, fuser)
for prefix, fuser in self.fusers.items()
if isinstance(fuser, MLAFuser)
}
if attn_cls is MLAAttention:
text_config._attn_implementation = "vllm_mla"
text_config._attn_implementation = VLLM_MLA_ATTN_IMPL
else:
# MLA model not using MLAAttention: recompute head_size for full attn
qk_nope_head_dim = getattr(text_config, "qk_nope_head_dim", 0)
@@ -570,12 +590,23 @@ class Base(
self.model_config.model_arch_config.head_size = qk_head_dim
logits_soft_cap = getattr(text_config, "attn_logit_softcapping", None)
layer_types = getattr(text_config, "layer_types", None)
pp_rank = self.pp_group.rank_in_group
pp_size = self.pp_group.world_size
start, end = get_pp_indices(text_config.num_hidden_layers, pp_rank, pp_size)
for i in range(start, end):
if i not in self.attention_fusers:
in_range = layer_types and i < len(layer_types)
layer = f"{i} ({layer_types[i]})" if in_range else str(i)
raise ValueError(
f"Layer {layer} does not dispatch through the Transformers "
"attention interface and vLLM has no other way to handle it."
)
prefix, attn_fuser = self.attention_fusers[i]
attn_module = self.get_submodule(prefix)
# `[i]` is the whole-model config unless the checkpoint is
# heterogeneous, in which case it is this layer's own geometry.
arch_config = self.model_config.model_arch_config[i]
@@ -583,8 +614,10 @@ class Base(
self.parallel_config, arch_config
)
head_size = arch_config.head_size
# Default to Llama scale, maybe updated in vllm_attention_forward
scale = head_size**-0.5
scale = attn_fuser.scale(attn_module)
# Default to Llama scale if AttentionFuser couldn't identify it
if scale is None:
scale = head_size**-0.5
num_kv_heads = self.model_config.get_num_kv_heads(
self.parallel_config, arch_config
)
@@ -598,17 +631,22 @@ class Base(
)
if attn_cls is MLAAttention:
prefix, fuser = mla_fusers[i]
mla_module = self.get_submodule(prefix)
mla_fuser = next(
(f for f in self.fusers[prefix] if isinstance(f, MLAFuser)), None
)
if mla_fuser is None:
raise ValueError(
f"Layer {i} ({prefix}) did not fuse as MLA, but the rest of "
"the model did, so its latent KV cache cannot be allocated."
)
dims = get_mla_dims(self.model_config)
kwargs.update(
scale=mla_module.scaling,
qk_nope_head_dim=dims.qk_nope_head_dim,
qk_rope_head_dim=dims.qk_rope_head_dim,
v_head_dim=dims.v_head_dim,
q_lora_rank=dims.q_lora_rank,
kv_lora_rank=dims.kv_lora_rank,
kv_b_proj=mla_module.get_submodule(fuser.kv_b_proj_name),
kv_b_proj=attn_module.get_submodule(mla_fuser.kv_b_proj_name),
)
else:
kwargs.update(
@@ -618,17 +656,20 @@ class Base(
)
# Handle interleaved sliding window attention
if (
hasattr(text_config, "layer_types")
and text_config.layer_types[i] == "sliding_attention"
):
if layer_types and layer_types[i] == "sliding_attention":
kwargs["per_layer_sliding_window"] = text_config.sliding_window
attn_instance = attn_cls(**kwargs)
if attn_cls is MLAAttention:
# Attach MLA attn_instance to mla_module so it appears in
# model.named_modules() and runs its process_weights_after_loading
mla_module._vllm_mla_attn = attn_instance
# Attach the instance to the module it serves, which is the only way it
# is in the module tree at all: a plain dict is invisible to
# `nn.Module.__setattr__`. Being in the tree runs its
# `process_weights_after_loading` and moves its KV-scale buffers with
# `model.to(device)`, and lets the attention interface read it off the
# module rather than index by `module.layer_idx`, an `int` that bakes
# layer identity into the traced graph and so costs one compiled
# artifact per layer.
setattr(attn_module, VLLM_ATTN_ATTR, attn_instance)
attention_instances[i] = attn_instance
return attention_instances
@@ -645,7 +686,11 @@ class Base(
return EncoderOnlyAttention
if self.model_config.use_mla:
self.check_version("5.15.0.dev0", "optimized MLA support")
if any(isinstance(fuser, MLAFuser) for fuser in self.fusers.values()):
if any(
isinstance(fuser, MLAFuser)
for fusers in self.fusers.values()
for fuser in fusers
):
return MLAAttention
logger.warning_once(
"This model uses MLA but `MLAFuser` failed to match and/or fuse any "
@@ -731,7 +776,6 @@ class Base(
inputs_embeds=inputs_embeds,
use_cache=False,
position_ids=positions,
attention_instances=self.attention_instances,
return_dict=False,
**self._output_aux_hidden_states_kwargs,
**kwargs,