change setup version code and Add support for Z100/Z100L

install_dcu.sh

 #!/bin/bash
 set -e
 
-# clear build dirs
+# 清理构建目录和旧的分发文件
 rm -rf build
+rm -rf dist
 rm -rf *.egg-info
 rm -rf ktransformers/ktransformers_ext/build
 rm -rf ktransformers/ktransformers_ext/cuda/build
 rm -rf ktransformers/ktransformers_ext/cuda/dist
 rm -rf ktransformers/ktransformers_ext/cuda/*.egg-info
-
-echo "Installing python dependencies from requirements.txt"
-pip install -r requirements-local_chat.txt
+echo "初始化Git子模块..."
+git submodule update --init --recursive
 
 export USE_FASTPT_CUDA=True
 export CMAKE_BUILD_PARALLEL_LEVEL=32
 
-echo "Installing ktransformers"
-KTRANSFORMERS_FORCE_BUILD=TRUE pip install . --no-build-isolation
-echo "Installation completed successfully"
+echo "构建ktransformers wheel包"
+
+mkdir -p dist
+
+KTRANSFORMERS_FORCE_BUILD=TRUE pip wheel . -w dist --no-build-isolation --no-deps
+
+echo "生成的wheel包位于："
+ls -l dist/*.whl
+
+echo "构建成功！wheel包已生成在dist目录"
+

ktransformers/__init__.py

@@ -8,4 +8,5 @@ Version      : 1.0.0
 LastEditors  : chenxl
 LastEditTime : 2025-02-15 03:53:02
 '''
-__version__ = "0.2.3.post1"
+__version__ = "0.2.3post1"
+__hcu_version__ = '0.2.3post1+das.dtk2504'

ktransformers/local_chat.py

@@ -28,7 +28,8 @@ from ktransformers.models.modeling_qwen2_moe import Qwen2MoeForCausalLM
 from ktransformers.models.modeling_deepseek_v3 import DeepseekV3ForCausalLM
 from ktransformers.models.modeling_llama import LlamaForCausalLM
 from ktransformers.models.modeling_mixtral import MixtralForCausalLM
-from ktransformers.util.utils import prefill_and_generate, get_compute_capability
+#from ktransformers.util.utils import prefill_and_generate, get_compute_capability
+from ktransformers.util.utils import prefill_and_generate, get_compute_capability, get_device_name
 from ktransformers.server.config.config import Config
 from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled
 
@@ -169,7 +170,8 @@ def local_chat(
             assert Config().long_context_config['max_seq_len'] > input_tensor.shape[1] + max_new_tokens, \
             "please change max_seq_len in  ~/.ktransformers/config.yaml"
         
-        if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or config.architectures[0] == "DeepseekV3ForCausalLM") and flashinfer_enabled and get_compute_capability() >= 8:
+        #if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or config.architectures[0] == "DeepseekV3ForCausalLM") and flashinfer_enabled and get_compute_capability() >= 8:
+        if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or "DeepseekV3ForCausalLM") and flashinfer_enabled and (get_compute_capability() >= 8 or ("Z100" in get_device_name()) or ("Z100L" in get_device_name())):
             generated = prefill_and_generate(
                 model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph, mode = mode, force_think = force_think, chunk_prefill_size = chunk_prefill_size,
                 use_flashinfer_mla = True, num_heads = config.num_attention_heads, head_dim_ckv = config.kv_lora_rank, head_dim_kpe = config.qk_rope_head_dim, q_head_dim = config.qk_rope_head_dim + config.qk_nope_head_dim

ktransformers/operators/attention.py

@@ -16,7 +16,8 @@ from ktransformers.models.modeling_deepseek import DeepseekV2Attention, apply_ro
 from typing import Optional, Tuple
 from ktransformers.operators.base_operator import BaseInjectedModule
 from ktransformers.util.custom_gguf import GGUFLoader
-from ktransformers.util.utils import get_compute_capability
+#from ktransformers.util.utils import get_compute_capability
+from ktransformers.util.utils import get_compute_capability, get_device_name
 import logging
 from transformers.configuration_utils import PretrainedConfig
 from transformers.cache_utils import Cache
@@ -589,8 +590,10 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if os.name == 'nt' or get_compute_capability()<8:
-            print("for Windows or GPU before ampere, use forward_windows")
+        #if os.name == 'nt' or get_compute_capability()<8:
+            #print("for Windows or GPU before ampere, use forward_windows")
+        if os.name == 'nt' or get_compute_capability()<8 or ("Z100" in get_device_name()) or ("Z100L" in get_device_name()):
+            print("for Windows or GPU before ampere or Z100/Z100L, use forward_windows")
             return self.forward_windows(
                 hidden_states,
                 attention_mask,

ktransformers/operators/models.py

@@ -56,7 +56,8 @@ from ktransformers.models.modeling_deepseek import (
 from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
 from ktransformers.models.configuration_llama import LlamaConfig
 from ktransformers.operators.base_operator import BaseInjectedModule
-from ktransformers.util.utils import InferenceState, get_compute_capability
+#from ktransformers.util.utils import InferenceState, get_compute_capability
+from ktransformers.util.utils import InferenceState, get_compute_capability, get_device_name
 from ktransformers.util.custom_gguf import GGUFLoader
 from transformers.configuration_utils import PretrainedConfig
 from ktransformers.models.modeling_llama import (
@@ -649,8 +650,10 @@ class KDeepseekV2Model(BaseInjectedModule):
         if per_layer_prefill_flag:
             causal_mask = None
         else:
-            if os.name == 'nt' or get_compute_capability()<8:
-                print("for Windows or GPU before ampere, use forward_windows")
+            #if os.name == 'nt' or get_compute_capability()<8:
+             #   print("for Windows or GPU before ampere, use forward_windows")
+            if os.name == 'nt' or get_compute_capability()<8 or ("Z100" in get_device_name()) or ("Z100L" in get_device_name()):
+                print("for Windows or GPU before ampere or Z100/Z100L, use forward_windows")
                 # only use mask in forward windows or can't flash attn
                 causal_mask = self._update_causal_mask(
                     attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions

ktransformers/util/utils.py

@@ -33,6 +33,18 @@ def get_compute_capability(device:torch.device = None):
         else:
             return torch.cuda.get_device_properties(device)
 
+
+def get_device_name(device:torch.device = None):
+    if torch.cuda.is_available():
+        if device is None:
+            num_gpus = torch.cuda.device_count()
+            gpu_name = []
+            for gpu_id in range(num_gpus):
+                gpu_name.append(torch.cuda.get_device_name(gpu_id))
+            return gpu_name
+        else:
+            return torch.cuda.get_device_name(device)
+
 def set_module(model, submodule_key, module):
     tokens = submodule_key.split('.')
     sub_tokens = tokens[:-1]

setup.py

@@ -377,8 +377,49 @@ elif MUSA_HOME is not None:
 else:
     raise ValueError("Unsupported backend: CUDA_HOME and MUSA_HOME are not set.")
 
+ROCM_PATH = os.getenv('ROCM_PATH')
+dtk_path = ROCM_PATH + '/.info/rocm_version'
+with open(dtk_path, 'r') as file:
+    content = file.read().strip()
+dtk_version = content.replace('.', '')
+print(dtk_version)
+cwd = os.path.dirname(os.path.abspath(__file__))
+ver_path = os.path.join(cwd, "ktransformers", "__init__.py")
+with open(ver_path, "r", encoding="utf-8") as file:
+    for line in file:
+        match = re.search(r'^__version__\s*=\s*["\'](.*?)["\']', line)
+        if match:
+            k_version = match.group(1)
+            break
+    else:
+        raise RuntimeError("未找到 __version__ 信息")
+
+with open(ver_path, 'r') as f:
+    lines = f.readlines()
+
+# 检查是否存在 __hcu_version__
+found = False
+new_lines = []
+for line in lines:
+    if line.startswith("__hcu_version__"):
+        # 替换已有的 __hcu_version__
+        version = k_version + '+das.dtk' + dtk_version
+        new_lines.append(f"__hcu_version__ = '{version}'\n")
+        found = True
+    else:
+        new_lines.append(line)
+
+# 如果未找到 __hcu_version__，则追加到文件末尾
+if not found:
+    version = k_version + '+das.dtk' + dtk_version
+    new_lines.append(f"__hcu_version__ = '{version}'\n")
+
+# 写回文件
+with open(ver_path, 'w') as f:
+    f.writelines(new_lines)
+
 setup(
-    version=VersionInfo().get_package_version(),
+    version=k_version + '+das.dtk' + dtk_version,
     cmdclass={"bdist_wheel":BuildWheelsCommand ,"build_ext": CMakeBuild},
     ext_modules=[
         CMakeExtension("cpuinfer_ext"),