[XPU] Misc changes to support XPU (#39817)
Signed-off-by: P V R K Jyothendra Varma <polisetty.v.r.k.jyothendra.varma@intel.com>
This commit is contained in:
@@ -223,6 +223,7 @@ class ExecKernel(msgspec.Struct):
|
||||
"triton",
|
||||
"aiter",
|
||||
"trtllm",
|
||||
"intel_xpu",
|
||||
],
|
||||
resolvable=True,
|
||||
),
|
||||
@@ -263,6 +264,7 @@ class ExecKernel(msgspec.Struct):
|
||||
"triton",
|
||||
"aiter",
|
||||
"trtllm",
|
||||
"intel_xpu",
|
||||
],
|
||||
resolvable=True,
|
||||
),
|
||||
|
||||
@@ -586,11 +586,6 @@ def _dsa_split_backend_resolution(view: Any) -> dict:
|
||||
declared["dsa_prefill_backend"] = "intel_xpu"
|
||||
if view.dsa_decode_backend is None:
|
||||
declared["dsa_decode_backend"] = "intel_xpu"
|
||||
# sgl-kernel topk ops (the default) are CUDA-only; fall back to the
|
||||
# torch-native topk implementation on XPU, unless the user already
|
||||
# picked a different backend explicitly (e.g. "flashinfer").
|
||||
if view.dsa_topk_backend == "sgl-kernel":
|
||||
declared["dsa_topk_backend"] = "torch"
|
||||
logger.warning(
|
||||
"Set DSA backends for XPU: prefill=%s, decode=%s, topk=%s.",
|
||||
declared.get("dsa_prefill_backend", view.dsa_prefill_backend),
|
||||
|
||||
@@ -1274,8 +1274,10 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, BaseFusedOp):
|
||||
b1=getattr(layer, "w13_weight_bias", None),
|
||||
b2=getattr(layer, "w2_weight_bias", None),
|
||||
activation=moe_runner_config.activation,
|
||||
routed_scaling_factor=moe_runner_config.routed_scaling_factor,
|
||||
gemm1_alpha=moe_runner_config.gemm1_alpha,
|
||||
gemm1_limit=moe_runner_config.gemm1_clamp_limit,
|
||||
swiglu_limit=moe_runner_config.swiglu_limit,
|
||||
)
|
||||
return StandardCombineInput(hidden_states=output)
|
||||
else:
|
||||
|
||||
@@ -154,8 +154,8 @@ class HYV3ForCausalLMNextN(nn.Module):
|
||||
del self.lm_head.weight
|
||||
self.model.embed_tokens.weight = embed
|
||||
self.lm_head.weight = head
|
||||
torch.cuda.empty_cache()
|
||||
torch.cuda.synchronize()
|
||||
torch.get_device_module().empty_cache()
|
||||
torch.get_device_module().synchronize()
|
||||
|
||||
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
||||
nextn_layer_id = self.config.num_hidden_layers
|
||||
|
||||
Reference in New Issue
Block a user