Skip to content

[Bugfix] Fix topk_ids indices_type for CUTLASS w8a8 FP8 MoE #20166

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Jul 8, 2025
Merged

[Bugfix] Fix topk_ids indices_type for CUTLASS w8a8 FP8 MoE #20166

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
6d0e19a
[Bugfix] Fix topk_ids indices_type for cutlass w8a8 fp8 moe
minosfuture Jun 24, 2025
7c57bb0
Address comment: change moe_data topk_ids type back to int32_t
minosfuture Jun 27, 2025
376bcde
Address comment: add expert_map==None assertion in pplx_prepare_finalize
minosfuture Jul 3, 2025
6ca83a4
Fix after rebase: cutlass_moe_fp8 signature is changed
minosfuture Jul 3, 2025
ec668c3
Address comment: change pplx id type to int32
minosfuture Jul 3, 2025
beeba9a
Address comment: revert type change for deepep_ll for now
minosfuture Jul 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions csrc/quantization/cutlass_w8a8/moe/moe_data.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@

constexpr uint64_t THREADS_PER_EXPERT = 512;

__global__ void compute_problem_sizes(const uint32_t* __restrict__ topk_ids,
__global__ void compute_problem_sizes(const int32_t* __restrict__ topk_ids,
int32_t* problem_sizes1,
int32_t* problem_sizes2,
int32_t* atomic_buffer,
Expand Down Expand Up @@ -62,7 +62,7 @@ __global__ void compute_expert_blockscale_offsets(
}
}

__global__ void compute_arg_sorts(const uint32_t* __restrict__ topk_ids,
__global__ void compute_arg_sorts(const int32_t* __restrict__ topk_ids,
const int32_t* __restrict__ expert_offsets,
int32_t* input_permutation,
int32_t* output_permutation,
Expand Down Expand Up @@ -103,7 +103,7 @@ void get_cutlass_moe_mm_data_caller(

int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
compute_problem_sizes<<<num_experts, num_threads, 0, stream>>>(
static_cast<const uint32_t*>(topk_ids.data_ptr()),
static_cast<const int32_t*>(topk_ids.data_ptr()),
static_cast<int32_t*>(problem_sizes1.data_ptr()),
static_cast<int32_t*>(problem_sizes2.data_ptr()),
static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n, k);
Expand All @@ -120,7 +120,7 @@ void get_cutlass_moe_mm_data_caller(
static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
}
compute_arg_sorts<<<num_experts, num_threads, 0, stream>>>(
static_cast<const uint32_t*>(topk_ids.data_ptr()),
static_cast<const int32_t*>(topk_ids.data_ptr()),
static_cast<const int32_t*>(expert_offsets.data_ptr()),
static_cast<int32_t*>(input_permutation.data_ptr()),
static_cast<int32_t*>(output_permutation.data_ptr()),
Expand Down
6 changes: 4 additions & 2 deletions vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,7 @@ def max_num_tokens_per_rank(self) -> Optional[int]:
return self.max_num_tokens

def topk_indices_dtype(self) -> Optional[torch.dtype]:
return torch.uint32
return torch.int32
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

note @tlrmchlsmth @varun-sundar-rabindranath - this appears to have broken the PPLX backend as the pplx dispatch function expects uint32

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more. 

VLLM_ALL2ALL_BACKEND="pplx" vllm serve Qwen/Qwen3-30B-A3B-FP8 --data-parallel-size 2 --enable-expert-parallel --enforce-eager

result:

(EngineCore_0 pid=1149186) ERROR 07-12 19:10:48 [core.py:586] RuntimeError: indices must be of type UInt32
(EngineCore_0 pid=1149186) Process EngineCore_0:
(EngineCore_0 pid=1149186) Traceback (most recent call last):
(EngineCore_0 pid=1149186)   File "/home/rshaw/.local/share/uv/python/cpython-3.12.11-linux-x86_64-gnu/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
(EngineCore_0 pid=1149186)     self.run()
(EngineCore_0 pid=1149186)   File "/home/rshaw/.local/share/uv/python/cpython-3.12.11-linux-x86_64-gnu/lib/python3.12/multiprocessing/process.py", line 108, in run
(EngineCore_0 pid=1149186)     self._target(*self._args, **self._kwargs)
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/engine/core.py", line 590, in run_engine_core
(EngineCore_0 pid=1149186)     raise e
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/engine/core.py", line 575, in run_engine_core
(EngineCore_0 pid=1149186)     engine_core = DPEngineCoreProc(*args, **kwargs)
(EngineCore_0 pid=1149186)                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/engine/core.py", line 835, in __init__
(EngineCore_0 pid=1149186)     super().__init__(vllm_config, local_client, handshake_address,
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/engine/core.py", line 404, in __init__
(EngineCore_0 pid=1149186)     super().__init__(vllm_config, executor_class, log_stats,
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/engine/core.py", line 82, in __init__
(EngineCore_0 pid=1149186)     self._initialize_kv_caches(vllm_config)
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/engine/core.py", line 142, in _initialize_kv_caches
(EngineCore_0 pid=1149186)     available_gpu_memory = self.model_executor.determine_available_memory()
(EngineCore_0 pid=1149186)                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/executor/abstract.py", line 76, in determine_available_memory
(EngineCore_0 pid=1149186)     output = self.collective_rpc("determine_available_memory")
(EngineCore_0 pid=1149186)              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/executor/uniproc_executor.py", line 57, in collective_rpc
(EngineCore_0 pid=1149186)     answer = run_method(self.driver_worker, method, args, kwargs)
(EngineCore_0 pid=1149186)              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/utils/__init__.py", line 2955, in run_method
(EngineCore_0 pid=1149186)     return func(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
(EngineCore_0 pid=1149186)     return func(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/worker/gpu_worker.py", line 219, in determine_available_memory
(EngineCore_0 pid=1149186)     self.model_runner.profile_run()
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/worker/gpu_model_runner.py", line 2239, in profile_run
(EngineCore_0 pid=1149186)     = self._dummy_run(self.max_num_tokens, is_profile=True)
(EngineCore_0 pid=1149186)       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
(EngineCore_0 pid=1149186)     return func(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/v1/worker/gpu_model_runner.py", line 2020, in _dummy_run
(EngineCore_0 pid=1149186)     outputs = model(
(EngineCore_0 pid=1149186)               ^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
(EngineCore_0 pid=1149186)     return self._call_impl(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
(EngineCore_0 pid=1149186)     return forward_call(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/models/qwen3_moe.py", line 529, in forward
(EngineCore_0 pid=1149186)     hidden_states = self.model(input_ids, positions, intermediate_tensors,
(EngineCore_0 pid=1149186)                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/compilation/decorators.py", line 173, in __call__
(EngineCore_0 pid=1149186)     return self.forward(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/models/qwen3_moe.py", line 369, in forward
(EngineCore_0 pid=1149186)     hidden_states, residual = layer(positions, hidden_states, residual)
(EngineCore_0 pid=1149186)                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
(EngineCore_0 pid=1149186)     return self._call_impl(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
(EngineCore_0 pid=1149186)     return forward_call(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/models/qwen3_moe.py", line 313, in forward
(EngineCore_0 pid=1149186)     hidden_states = self.mlp(hidden_states)
(EngineCore_0 pid=1149186)                     ^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
(EngineCore_0 pid=1149186)     return self._call_impl(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
(EngineCore_0 pid=1149186)     return forward_call(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/models/qwen3_moe.py", line 136, in forward
(EngineCore_0 pid=1149186)     final_hidden_states = self.experts(hidden_states=hidden_states,
(EngineCore_0 pid=1149186)                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
(EngineCore_0 pid=1149186)     return self._call_impl(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
(EngineCore_0 pid=1149186)     return forward_call(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/layer.py", line 1381, in forward
(EngineCore_0 pid=1149186)     return torch.ops.vllm.moe_forward(hidden_states, router_logits,
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/_ops.py", line 1158, in __call__
(EngineCore_0 pid=1149186)     return self._op(*args, **(kwargs or {}))
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/layer.py", line 1566, in moe_forward
(EngineCore_0 pid=1149186)     return self.forward_impl(hidden_states, router_logits)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/layer.py", line 1465, in forward_impl
(EngineCore_0 pid=1149186)     return self.forward_impl_chunked(hidden_states, router_logits)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/layer.py", line 1454, in forward_impl_chunked
(EngineCore_0 pid=1149186)     process_chunk(chunk_start,
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/layer.py", line 1415, in process_chunk
(EngineCore_0 pid=1149186)     final_hidden_states = self.quant_method.apply(
(EngineCore_0 pid=1149186)                           ^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/quantization/fp8.py", line 970, in apply
(EngineCore_0 pid=1149186)     return self.fused_experts(
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
(EngineCore_0 pid=1149186)     return self._call_impl(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
(EngineCore_0 pid=1149186)     return forward_call(*args, **kwargs)
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/modular_kernel.py", line 689, in forward
(EngineCore_0 pid=1149186)     _expert_topk_weights) = self.prepare_finalize.prepare(
(EngineCore_0 pid=1149186)                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py", line 202, in prepare
(EngineCore_0 pid=1149186)     self.a2a.dispatch(
(EngineCore_0 pid=1149186)   File "/home/rshaw/vllm/tools/ep_kernels/ep_kernels_workspace/pplx-kernels/src/pplx_kernels/all_to_all.py", line 48, in dispatch
(EngineCore_0 pid=1149186)     self._dispatch_fn(
(EngineCore_0 pid=1149186)   File "/home/rshaw/.vllm/lib/python3.12/site-packages/torch/_ops.py", line 1158, in __call__
(EngineCore_0 pid=1149186)     return self._op(*args, **(kwargs or {}))
(EngineCore_0 pid=1149186)            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(EngineCore_0 pid=1149186) RuntimeError: indices must be of type UInt32

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@robertgshaw2-redhat that will be fixed by #20825

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

okay cool - just a note: I was using triton+pplx in this case

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry for the regression! I'll make sure to test it next time. We are also working on improving CI coverage and preparing a script to cover major models for easy local test.

VLLM_ALL2ALL_BACKEND="pplx" vllm serve Qwen/Qwen3-30B-A3B-FP8 --data-parallel-size 2 --enable-expert-parallel --enforce-eager


def prepare(
self,
Expand All @@ -100,7 +100,9 @@ def prepare(
hidden_dim = a1.size(-1) # K

assert topk_ids.size(0) == num_tokens
# assert expert_map is None, "NYI"
assert expert_map is None, """with expert map, -1 id is used for
non-local token; this causes error when casting ids to the
topk_indices_dtype() uint32"""
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added assertion and changed the id type here. test_pplx_moe.py passes. Let me know if we should run a model e2e and how. Thx! @tylertitsworth @bnellnm cc @yeqcharlotte

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking at this again, I don't think we need both the assertion and int32 change?
I don't know enough internals of pplx comms to decide this. Let me know how we wanna finalize this change, @tlrmchlsmth @bnellnm. Thanks!

Copy link
Contributor

@varun-sundar-rabindranath varun-sundar-rabindranath Jul 9, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @minosfuture Could we lift this assert ? I think PrepareAndFinalize implementations should allow callers to pass in a "valid" expert_map.

I believe a better solution is to make expert_map=None inside this function with comment with expert map, -1 id is used for non-local token; this causes error when casting ids to the topk_indices_dtype() uint32 to prevent issues with pplx + expert_map.

fyi @tlrmchlsmth

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

let me raise a PR for this! thanks for the catch.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

would it be problematic to clear expert_map assigning None to it?
do you know why an expert_map is passed in in the first place?

Copy link
Contributor

@varun-sundar-rabindranath varun-sundar-rabindranath Jul 9, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

would it be problematic to clear expert_map assigning None to it?

The function doesn't use the expert_map - so overriding it to None should be fine as it'd prevent any incorrect use.

do you know why an expert_map is passed in in the first place?

I actually dont see it being used in any of the implementations. I think it is okay to remove the argument altogether.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

parameter is needed for this is an inherited function.

I commented it all out. This essentially reverts the changes here. As a followup, I think we should fix where it passes an expert_map, and add the assertion back.

#20714 @varun-sundar-rabindranath @tlrmchlsmth pls help approve this fix pr and auto-merge it. thx!

Copy link
Contributor

@varun-sundar-rabindranath varun-sundar-rabindranath Jul 10, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

parameter is needed for this is an inherited function.

I was suggesting we could remove the var in the base-class as I don't see any implementation using it.
I approved it, we can do the cleanups in a followup. Thanks 🙌


# Is this always going to be a1.device?
device = a1.device
Expand Down
14 changes: 9 additions & 5 deletions vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -883,23 +883,27 @@ def apply(
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
)
e_score_correction_bias=e_score_correction_bias)

a1_scale = layer.w13_input_scale
a2_scale = layer.w2_input_scale
per_act_token = a1_scale.numel() != 1 if a1_scale is not None else (
a2_scale.numel() != 1 if a2_scale is not None else False)
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this is a fix needed after rebase #19636. cc @bnellnm @luccafong


return self.fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights,
topk_ids,
per_act_token=per_act_token,
activation=activation,
global_num_experts=global_num_experts,
expert_map=None if self.disable_expert_map else 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,
a1_scale=a1_scale,
a2_scale=a2_scale,
)


Expand Down