dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

fix(build): 回退qwen3_6_scripts+ex_engine到26e6cb40(能得分版本)

Browse Source 
唯一改动: computility-run.yaml max_model_len 80000→100000

26e6cb40是Sub520能在竞赛平台docker build成功并得分的版本
之后所有commit都导致docker build失败
根因: 新增的65个文件(vendor_overrides/prebuilt/*.so/wheels等)
可能触发了竞赛平台docker build的某个限制

本次回退:
- qwen3_6_scripts/: 110→45文件(删掉65个新增文件)
- ex_engine/: 恢复到26e6cb40完全一致
- Dockerfile: 恢复5个RUN步骤结构(已验证能build)
- computility-run.yaml: max_model_len=100000(避免replay 400拒绝)
This commit is contained in:
Claude
2026-08-12 01:33:24 +00:00
parent f8e8b6fb28
commit cf1b701afe
138 changed files with 10282 additions and 31623 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
ex_engine/csrc/ilu/ixformer.h
View File

@@ -34,9 +34,9 @@ torch::Tensor ixinfer_flash_attn_unpad_with_block_tables(
double scale,
double softcap,
bool sqrt_alibi,
const c10::optional<torch::Tensor>& alibi_slopes,
const c10::optional<torch::Tensor>& sinks,
c10::optional<torch::Tensor>& lse);
const std::optional<torch::Tensor>& alibi_slopes,
const std::optional<torch::Tensor>& sinks,
std::optional<torch::Tensor>& lse);
void silu_and_mul(torch::Tensor& input, torch::Tensor& output);
@@ -51,21 +51,21 @@ torch::Tensor xllm_paged_attention(
torch::Tensor& context_lens,
int64_t block_size,
int64_t max_context_len,
const c10::optional<torch::Tensor>& alibi_slopes,
const std::optional<torch::Tensor>& alibi_slopes,
bool causal,
int32_t window_left,
int32_t window_right,
double softcap,
bool enable_cuda_graph,
bool use_sqrt_alibi,
const c10::optional<torch::Tensor>& sinks);
const std::optional<torch::Tensor>& sinks);
torch::Tensor ixformer_linear(torch::Tensor& input,
torch::Tensor& weight,
int64_t act_type,
const c10::optional<torch::Tensor>& bias,
const c10::optional<torch::Tensor>& out,
const c10::optional<bool> persistent);
const std::optional<torch::Tensor>& bias,
const std::optional<torch::Tensor>& out,
const std::optional<bool> persistent);
torch::Tensor ixformer_linear_ex(torch::Tensor& input,
torch::Tensor& weight,
@@ -92,7 +92,7 @@ void residual_rms_norm(torch::Tensor& input,
torch::Tensor& weight,
torch::Tensor& output,
torch::Tensor& residual_output,
const c10::optional<torch::Tensor>& fused_bias,
const std::optional<torch::Tensor>& fused_bias,
double alpha,
double eps,
bool is_post);
@@ -100,7 +100,7 @@ void residual_rms_norm(torch::Tensor& input,
void rms_norm(torch::Tensor& input,
torch::Tensor& weight,
torch::Tensor& output,
const c10::optional<torch::Tensor>& fused_bias,
const std::optional<torch::Tensor>& fused_bias,
double eps);
void topk_softmax(torch::Tensor& topk_weights,