EngineX/xc-llm-kunlun
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

remove qwen2.py llama.py fix llama output

Browse Source
This commit is contained in:
hanhaowen
2025-12-31 11:31:26 +08:00
parent b3c30a3cb9
commit b015bb76fd
11 changed files with 65 additions and 1263 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm_kunlun/ops/__init__.py
View File

@@ -19,4 +19,5 @@ import vllm_kunlun.ops.rotary_embedding
import vllm_kunlun.ops.layernorm
import vllm_kunlun.ops.quantization.awq
import vllm_kunlun.ops.quantization.gptq
import vllm_kunlun.ops.vocab_parallel_embedding
import vllm_kunlun.ops.vocab_parallel_embedding
import vllm_kunlun.ops.linear

4
vllm_kunlun/ops/_kunlun_ops.py
View File

@@ -491,7 +491,7 @@ class KunlunOps:
d = y.shape[-1] // 2
output_shape = (y.shape[:-1] + (d, ))
out1 = torch.empty(output_shape, dtype=y.dtype, device=y.device)
torch.ops._C.swiglu(y, out1)
torch.ops._C.silu_and_mul(out1, y)
out = torch.empty(M,moe_top_k,
w2.shape[1],
@@ -570,7 +570,7 @@ class KunlunOps:
cur_token = repeat_x[selected_token]
up_gate = torch.empty(selected_token.sum(), up_gate_size//2,
dtype=cur_token.dtype, device=cur_token.device)
torch.ops._C.swiglu(cur_token@ w13_weight[i].T, up_gate)
torch.ops._C.silu_and_mul(up_gate, cur_token@ w13_weight[i].T)
out[selected_token] = up_gate @ w2_weight[i].T
output = (out.view(batch, top_k, hidden_size) * topk_weights.unsqueeze(2)).sum(dim=1).to(x.dtype)

2
vllm_kunlun/ops/activation.py
View File

@@ -98,7 +98,7 @@ class SiluAndMul(CustomOp):
d = x.shape[-1] // 2
output_shape = (x.shape[:-1] + (d, ))
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
torch.ops._C.swiglu(x, out)
torch.ops._C.silu_and_mul(out, x)
return out
def forward_kunlun(self, x: torch.Tensor) -> torch.Tensor: