Upload 6 files

.gitattributes

@@ -110,3 +110,4 @@ qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_p128_l7_together.axmodel filter=lfs diff=lf
 qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_p128_l8_together.axmodel filter=lfs diff=lfs merge=lfs -text
 qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_p128_l9_together.axmodel filter=lfs diff=lfs merge=lfs -text
 qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_post.axmodel filter=lfs diff=lfs merge=lfs -text
+main_ax650 filter=lfs diff=lfs merge=lfs -text

main_ax650

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d647c00b7f4f2d39216575e228e7713ef3efb029827d2ab90da733c686867788
+size 985272

main_axcl_x86

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6cba0be8df523e351789fcfa745772423096f52a3b0a760f8d8c9f5b8bb2ec82
-size 1022384
+oid sha256:d427677213e3b50b22ed55a1d0af10223f5b0711c1bf050d0655aedd3ffc8df9
+size 1839824

post_config.json

@@ -1,5 +1,5 @@
 {
-    "enable_temperature" : true,
+    "enable_temperature" : false,
     "temperature" : 0.9,
     
     "enable_repetition_penalty" : false,
@@ -10,5 +10,5 @@
     "top_p" : 0.8,
 
     "enable_top_k_sampling" : true,
-    "top_k" : 10
+    "top_k" : 1
 }

qwen2.5_tokenizer_uid.py

@@ -0,0 +1,199 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+import uuid
+
+# 全局字典：存储 uid 到 Tokenizer_Http 实例的映射
+tokenizers = {}
+
+class Tokenizer_Http():
+    def __init__(self):
+        model_id = "qwen2.5_tokenizer"
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+        self.messages = [
+            {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
+        ]
+        self.token_ids = []
+        
+        self.token_ids_cache = []
+
+    def encode(self, prompt, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+            text = self.tokenizer.apply_chat_template(
+                self.messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+            # print("生成的文本:\n============\n", text, "============\n")
+            self.token_ids = self.tokenizer.encode(text)[:-3]
+        self.messages.append({"role": "user", "content": prompt})
+        
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print("生成的文本:\n============\n", text, "============\n")
+        token_ids = self.tokenizer.encode(text)
+        # 找出新增部分
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        print(self.decode(diff))
+        return token_ids, diff
+
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+        
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+    
+    def reset(self, system_prompt="You are Qwen, created by Alibaba Cloud. You are a helpful assistant."):
+        self.messages = [
+            {"role": "system", "content": system_prompt},
+        ]
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)[:-3]
+        self.token_ids = token_ids
+        print(self.decode(token_ids))
+        return token_ids
+        
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print("GET 请求路径:", self.path)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        # 新增接口：获取 uid
+        if '/get_uid' in self.path:
+            new_uid = str(uuid.uuid4())
+            print("新 uid:", new_uid)
+            # 为该 uid 创建一个新的 Tokenizer_Http 实例
+            tokenizers[new_uid] = Tokenizer_Http()
+            msg = json.dumps({'uid': new_uid})
+        elif '/bos_id' in self.path:
+            # 获取 uid 参数（例如 ?uid=xxx）
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                bos_id = instance.bos_id
+                msg = json.dumps({'bos_id': bos_id if bos_id is not None else -1})
+        elif '/eos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                eos_id = instance.eos_id
+                msg = json.dumps({'eos_id': eos_id if eos_id is not None else -1})
+        else:
+            msg = json.dumps({'error': 'Invalid GET endpoint'})
+
+        print("响应消息:", msg)
+        self.wfile.write(msg.encode())
+
+    def do_POST(self):
+        content_length = int(self.headers.get('content-length', 0))
+        data = self.rfile.read(content_length).decode()
+        print("POST 请求路径:", self.path)
+        print("接收到的数据:", data)
+        req = json.loads(data)
+
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        if '/encode' in self.path:
+            # 请求数据中必须包含 uid, text, 和可选的 last_reply
+            uid = req.get('uid')
+            prompt = req.get('text')
+            last_reply = req.get('last_reply')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                token_ids, diff = instance.encode(prompt, last_reply)
+                msg = json.dumps({'token_ids': token_ids, 'diff': diff})
+        elif '/decode' in self.path:
+            uid = req.get('uid')
+            token_ids = req.get('token_ids')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                text = instance.decode(token_ids)
+                msg = json.dumps({'text': text})
+        elif '/reset' in self.path:
+            uid = req.get("uid")
+            system_prompt = req.get("system_prompt")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if system_prompt is not None:
+                    print("system_prompt:", system_prompt)
+                    token_ids = instance.reset(system_prompt)
+                    msg = json.dumps({'token_ids': token_ids})
+                else:
+                    token_ids = instance.reset()
+                    msg = json.dumps({'token_ids': token_ids})
+        else:
+            msg = json.dumps({'error': 'Invalid POST endpoint'})
+
+        print("响应消息:", msg)
+        self.wfile.write(msg.encode())
+
+    def get_query_param(self, key):
+        """
+        辅助函数：从 GET 请求的 URL 中获取查询参数的值
+        例如：/bos_id?uid=xxx
+        """
+        from urllib.parse import urlparse, parse_qs
+        query = urlparse(self.path).query
+        params = parse_qs(query)
+        values = params.get(key)
+        return values[0] if values else None
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--host', type=str, default='0.0.0.0')
+    parser.add_argument('--port', type=int, default=12345)
+    args = parser.parse_args()
+
+    host = (args.host, args.port)
+    print('Server running at http://%s:%s' % host)
+    server = HTTPServer(host, Request)
+    server.serve_forever()

run_qwen2.5_3b_gptq_int4_ctx_ax650.sh

@@ -0,0 +1,14 @@
+./main_ax650 \
+--system_prompt "You are Qwen, created by Alibaba Cloud. You are a helpful assistant." \
+--template_filename_axmodel "qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_p128_l%d_together.axmodel" \
+--axmodel_num 36 \
+--tokenizer_type 2 \
+--url_tokenizer_model "http://127.0.0.1:12345" \
+--filename_post_axmodel "qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_post.axmodel" \
+--filename_tokens_embed "qwen2.5-3b-gptq-int4-ctx-ax650/model.embed_tokens.weight.bfloat16.bin" \
+--tokens_embed_num 151936 \
+--tokens_embed_size 2048 \
+--use_mmap_load_embed 1 \
+--live_print 1
+
+# --kvcache_path /home/axera/ax-llm/build/kvcache_yuanqi \

run_qwen2.5_3b_gptq_int4_ctx_axcl_x86.sh

@@ -0,0 +1,15 @@
+./main_axcl_x86 \
+--system_prompt "You are Qwen, created by Alibaba Cloud. You are a helpful assistant." \
+--template_filename_axmodel "qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_p128_l%d_together.axmodel" \
+--axmodel_num 36 \
+--tokenizer_type 2 \
+--url_tokenizer_model "http://127.0.0.1:12345" \
+--filename_post_axmodel "qwen2.5-3b-gptq-int4-ctx-ax650/qwen2_post.axmodel" \
+--filename_tokens_embed "qwen2.5-3b-gptq-int4-ctx-ax650/model.embed_tokens.weight.bfloat16.bin" \
+--tokens_embed_num 151936 \
+--tokens_embed_size 2048 \
+--use_mmap_load_embed 0 \
+--live_print 1 \
+--devices 0
+
+# --kvcache_path /home/axera/ax-llm/build/kvcache_yuanqi \