biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Upstream Llama4 Support to Main (#16113)

Browse Source 
Signed-off-by: Aston Zhang <22279212+astonzhang@users.noreply.github.com>
Signed-off-by: Chris Thi <chris.c.thi@gmail.com>
Signed-off-by: drisspg <drisspguessous@gmail.com>
Signed-off-by: Jon Swenson <jmswen@gmail.com>
Signed-off-by: Keyun Tong <tongkeyun@gmail.com>
Signed-off-by: Lu Fang <fanglu@meta.com>
Signed-off-by: Xiaodong Wang <xdwang@meta.com>
Signed-off-by: Yang Chen <yangche@fb.com>
Signed-off-by: Ye (Charlotte) Qi <yeq@meta.com>
Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
Signed-off-by: Zijing Liu <liuzijing2014@gmail.com>
Signed-off-by: Lu Fang <lufang@fb.com>
Signed-off-by: Lu Fang <fanglu@fb.com>
Signed-off-by: Lucia Fang <fanglu@fb.com>
Signed-off-by: Roger Wang <ywang@roblox.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: Lu Fang <fanglu@fb.com>
Co-authored-by: Roger Wang <ywang@roblox.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Lu Fang
2025-04-07 08:06:27 -07:00
committed by GitHub
parent 8017c8db7f
commit 55dcce91df
43 changed files with 2436 additions and 155 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/layers/quantization/awq_marlin.py
View File

@@ -469,6 +469,7 @@ class AWQMoEMethod(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
assert activation == "silu", "Only SiLU activation is supported."
@@ -476,6 +477,10 @@ class AWQMoEMethod(FusedMoEMethodBase):
raise NotImplementedError(
"Expert Parallelism is not supported for "
"fused Marlin MoE method.")
if apply_router_weight_on_input:
raise NotImplementedError(
"Apply router weight on input is not supported for"
"fused Marlin MoE method.")
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,

38
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
View File

@@ -224,6 +224,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -240,20 +241,22 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
return fused_experts(x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=activation,
use_fp8_w8a8=True,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale)
return fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=activation,
apply_router_weight_on_input=apply_router_weight_on_input,
use_fp8_w8a8=True,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale)
class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
@@ -438,6 +441,7 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
@@ -474,6 +478,7 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
out_dtype=x.dtype,
apply_router_weight_on_input=apply_router_weight_on_input,
)
@@ -778,6 +783,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
assert activation == "silu", "Only SiLU activation is supported."
@@ -785,6 +791,10 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
raise NotImplementedError(
"Expert Parallelism is not supported for "
"fused Marlin MoE method.")
if apply_router_weight_on_input:
raise NotImplementedError(
"Apply router weight on input is not supported for "
"fused Marlin MoE method.")
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,

27
vllm/model_executor/layers/quantization/experts_int8.py
View File

@@ -113,6 +113,7 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -129,18 +130,20 @@ class ExpertsInt8MoEMethod(FusedMoEMethodBase):
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
return fused_experts(x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=activation,
use_int8_w8a16=True,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=layer.w13_scale,
w2_scale=layer.w2_scale)
return fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=activation,
use_int8_w8a16=True,
global_num_experts=global_num_experts,
apply_router_weight_on_input=apply_router_weight_on_input,
expert_map=expert_map,
w1_scale=layer.w13_scale,
w2_scale=layer.w2_scale)
@staticmethod
def quantizing_weight_loader(layer, weight_loader):

2
vllm/model_executor/layers/quantization/fp8.py
View File

@@ -773,6 +773,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -800,6 +801,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
activation=activation,
use_fp8_w8a8=True,
global_num_experts=global_num_experts,
apply_router_weight_on_input=apply_router_weight_on_input,
expert_map=expert_map,
w1_scale=(layer.w13_weight_scale_inv
if self.block_quant else layer.w13_weight_scale),

6
vllm/model_executor/layers/quantization/gguf.py
View File

@@ -338,9 +338,15 @@ class GGUFMoEMethod(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
):
assert activation == "silu", "Only SiLU activation is supported."
if apply_router_weight_on_input:
raise NotImplementedError(
"Apply router weight on input is not supported for"
"fused GGUF MoE method.")
topk_weights, topk_ids = FusedMoE.select_experts(
hidden_states=x,
router_logits=router_logits,

5
vllm/model_executor/layers/quantization/gptq_marlin.py
View File

@@ -592,9 +592,14 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
assert activation == "silu", "Only SiLU activation is supported."
if apply_router_weight_on_input is not None:
raise NotImplementedError(
"Apply router weight on input is not supported for"
"fused Marlin MoE method.")
# The input must currently be float16
orig_dtype = x.dtype

33
vllm/model_executor/layers/quantization/moe_wna16.py
View File

@@ -293,6 +293,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -312,21 +313,23 @@ class MoeWNA16Method(FusedMoEMethodBase):
weight_bits = self.quant_config.weight_bits
has_zp = self.quant_config.has_zp
return fused_experts(x,
layer.w13_qweight,
layer.w2_qweight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
use_int4_w4a16=weight_bits == 4,
use_int8_w8a16=weight_bits == 8,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=layer.w13_scales,
w2_scale=layer.w2_scales,
w1_zp=layer.w13_qzeros if has_zp else None,
w2_zp=layer.w2_qzeros if has_zp else None,
block_shape=[0, layer.group_size])
return fused_experts(
x,
layer.w13_qweight,
layer.w2_qweight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
use_int4_w4a16=weight_bits == 4,
use_int8_w8a16=weight_bits == 8,
global_num_experts=global_num_experts,
apply_router_weight_on_input=apply_router_weight_on_input,
expert_map=expert_map,
w1_scale=layer.w13_scales,
w2_scale=layer.w2_scales,
w1_zp=layer.w13_qzeros if has_zp else None,
w2_zp=layer.w2_qzeros if has_zp else None,
block_shape=[0, layer.group_size])
@staticmethod
def get_weight_loader(layer, weight_loader):

30
vllm/model_executor/layers/quantization/quark/quark_moe.py
View File

@@ -202,6 +202,8 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
activation: str = "silu",
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_experts
@@ -217,16 +219,18 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
return fused_experts(x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
use_fp8_w8a8=True,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale)
return fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
use_fp8_w8a8=True,
global_num_experts=global_num_experts,
apply_router_weight_on_input=apply_router_weight_on_input,
expert_map=expert_map,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale)