0b0c47fddd
Sub 655 root causes (confirmed from log analysis): 1. protocol.py: max_completion_tokens never folded into max_tokens → 162/881 replay requests rejected 400 (extra_forbidden) 2. max_num_seqs=1 → t2_n_2 test fails (needs n=2) 3. max_model_len=131072 → OOM crash at 62% replay, opencompass all 0 Fixes: - protocol.py: model_validator fold_max_completion_tokens - yaml: max_num_seqs=2, max_model_len=80000, PYTORCH_CUDA_ALLOC_CONF - topk_softmax stays =0 (corex CUB BlockReduce incompatible on BI-V100) xllm_latest import to ex_engine/: - npu_torch layers: GDN(1164L), Qwen3.5 GDN, attention, fused_moe - cuda/moe kernels: topk_softmax_kernels.cuh, moe_combine, moe_compute_index - npu kernels: causal_conv1d, recurrent_gated_delta_rule - model headers: qwen3_5.h, qwen3_next.h
80 lines
2.5 KiB
C++
80 lines
2.5 KiB
C++
/* Copyright 2026 The xLLM Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
https://github.com/jd-opensource/xllm/blob/main/LICENSE
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
==============================================================================*/
|
|
|
|
#pragma once
|
|
|
|
#include <torch/torch.h>
|
|
|
|
#include "attention.h"
|
|
#include "framework/kv_cache/kv_cache.h"
|
|
#include "framework/model/model_args.h"
|
|
#include "framework/parallel_state/parallel_args.h"
|
|
#include "framework/quant_args.h"
|
|
#include "framework/state_dict/state_dict.h"
|
|
#include "layers/common/linear.h"
|
|
#include "layers/common/partial_rotary_embedding.h"
|
|
#include "layers/common/qwen3_next_rms_norm.h"
|
|
#include "layers/common/rotary_embedding.h"
|
|
|
|
namespace xllm {
|
|
namespace layer {
|
|
|
|
class Qwen3_5AttentionImpl : public torch::nn::Module {
|
|
public:
|
|
Qwen3_5AttentionImpl() = default;
|
|
Qwen3_5AttentionImpl(const ModelArgs& args,
|
|
const QuantArgs& quant_args,
|
|
const ParallelArgs& parallel_args,
|
|
const torch::TensorOptions& options,
|
|
int32_t layer_id);
|
|
|
|
torch::Tensor forward(const torch::Tensor& positions,
|
|
const torch::Tensor& hidden_states,
|
|
const AttentionMetadata& attn_metadata,
|
|
KVCache& kv_cache);
|
|
|
|
void load_state_dict(const StateDict& state_dict);
|
|
void rotary_emb_forward(torch::Tensor& q,
|
|
torch::Tensor& k,
|
|
const torch::Tensor& positions,
|
|
const AttentionMetadata& attn_metadata);
|
|
|
|
private:
|
|
int64_t num_heads_;
|
|
int64_t num_kv_heads_;
|
|
int64_t num_kv_head_replicas_;
|
|
int64_t head_dim_;
|
|
int64_t q_size_;
|
|
int64_t kv_size_;
|
|
float scaling_;
|
|
bool attn_output_gate_;
|
|
int32_t layer_id_;
|
|
int32_t rank_;
|
|
|
|
QKVParallelLinear qkv_proj_{nullptr};
|
|
RowParallelLinear o_proj_{nullptr};
|
|
|
|
Qwen3NextRMSNorm q_norm_{nullptr};
|
|
Qwen3NextRMSNorm k_norm_{nullptr};
|
|
|
|
Attention attn_{nullptr};
|
|
MRotaryEmbedding rotary_emb_{nullptr};
|
|
torch::Tensor mrope_cu_seq_lens_;
|
|
};
|
|
TORCH_MODULE(Qwen3_5Attention);
|
|
|
|
} // namespace layer
|
|
} // namespace xllm
|