Update config.production-modelscope.yaml
Browse files
config.production-modelscope.yaml
CHANGED
|
@@ -1,68 +1,68 @@
|
|
| 1 |
-
HOST: "0.0.0.0"
|
| 2 |
-
PORT: 7860
|
| 3 |
-
STRATEGY: "cuda fp16"
|
| 4 |
-
RWKV_CUDA_ON: True
|
| 5 |
-
CHUNK_LEN: 256
|
| 6 |
-
MODELS:
|
| 7 |
-
- SERVICE_NAME: "rwkv7-g1a-2.9b-20250924-ctx4096"
|
| 8 |
-
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-2.9b-20250924-ctx4096.pth"
|
| 9 |
-
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 10 |
-
REASONING: True
|
| 11 |
-
DEFAULT_SAMPLER:
|
| 12 |
-
max_tokens: 4096
|
| 13 |
-
temperature: 1
|
| 14 |
-
top_p: 0.3
|
| 15 |
-
presence_penalty: 0.5
|
| 16 |
-
count_penalty: 0.5
|
| 17 |
-
penalty_decay: 0.996
|
| 18 |
-
stop:
|
| 19 |
-
- "\n\n"
|
| 20 |
-
stop_tokens:
|
| 21 |
-
- 0
|
| 22 |
-
- SERVICE_NAME: "rwkv7-g1a-1.5b-20250922-ctx4096"
|
| 23 |
-
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-1.5b-20250922-ctx4096.pth"
|
| 24 |
-
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 25 |
-
REASONING: True
|
| 26 |
-
DEFAULT_SAMPLER:
|
| 27 |
-
max_tokens: 4096
|
| 28 |
-
temperature: 1
|
| 29 |
-
top_p: 0.3
|
| 30 |
-
presence_penalty: 0.5
|
| 31 |
-
count_penalty: 0.5
|
| 32 |
-
penalty_decay: 0.996
|
| 33 |
-
stop:
|
| 34 |
-
- "\n\n"
|
| 35 |
-
stop_tokens:
|
| 36 |
-
- 0
|
| 37 |
-
- SERVICE_NAME: "rwkv7-g1a-0.4b-20250905-ctx4096"
|
| 38 |
-
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-0.4b-20250905-ctx4096.pth"
|
| 39 |
-
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 40 |
-
REASONING: True
|
| 41 |
-
DEFAULT_SAMPLER:
|
| 42 |
-
max_tokens: 4096
|
| 43 |
-
temperature: 1
|
| 44 |
-
top_p: 0.3
|
| 45 |
-
presence_penalty: 0.5
|
| 46 |
-
count_penalty: 0.5
|
| 47 |
-
penalty_decay: 0.996
|
| 48 |
-
stop:
|
| 49 |
-
- "\n\n"
|
| 50 |
-
stop_tokens:
|
| 51 |
-
- 0
|
| 52 |
-
- SERVICE_NAME: "rwkv7-g1a-0.1b-20250728-ctx4096"
|
| 53 |
-
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-0.1b-20250728-ctx4096.pth"
|
| 54 |
-
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 55 |
-
REASONING: True
|
| 56 |
-
DEFAULT_CHAT: True
|
| 57 |
-
DEFAULT_REASONING: True
|
| 58 |
-
DEFAULT_SAMPLER:
|
| 59 |
-
max_tokens: 4096
|
| 60 |
-
temperature: 1
|
| 61 |
-
top_p: 0.3
|
| 62 |
-
presence_penalty: 0.5
|
| 63 |
-
count_penalty: 0.5
|
| 64 |
-
penalty_decay: 0.996
|
| 65 |
-
stop:
|
| 66 |
-
- "\n\n"
|
| 67 |
-
stop_tokens:
|
| 68 |
-
- 0
|
|
|
|
| 1 |
+
HOST: "0.0.0.0"
|
| 2 |
+
PORT: 7860
|
| 3 |
+
STRATEGY: "cuda fp16"
|
| 4 |
+
RWKV_CUDA_ON: True
|
| 5 |
+
CHUNK_LEN: 256
|
| 6 |
+
MODELS:
|
| 7 |
+
- SERVICE_NAME: "rwkv7-g1a-2.9b-20250924-ctx4096"
|
| 8 |
+
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-2.9b-20250924-ctx4096.pth"
|
| 9 |
+
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 10 |
+
REASONING: True
|
| 11 |
+
DEFAULT_SAMPLER:
|
| 12 |
+
max_tokens: 4096
|
| 13 |
+
temperature: 0.1
|
| 14 |
+
top_p: 0.3
|
| 15 |
+
presence_penalty: 0.5
|
| 16 |
+
count_penalty: 0.5
|
| 17 |
+
penalty_decay: 0.996
|
| 18 |
+
stop:
|
| 19 |
+
- "\n\n"
|
| 20 |
+
stop_tokens:
|
| 21 |
+
- 0
|
| 22 |
+
- SERVICE_NAME: "rwkv7-g1a-1.5b-20250922-ctx4096"
|
| 23 |
+
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-1.5b-20250922-ctx4096.pth"
|
| 24 |
+
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 25 |
+
REASONING: True
|
| 26 |
+
DEFAULT_SAMPLER:
|
| 27 |
+
max_tokens: 4096
|
| 28 |
+
temperature: 0.1
|
| 29 |
+
top_p: 0.3
|
| 30 |
+
presence_penalty: 0.5
|
| 31 |
+
count_penalty: 0.5
|
| 32 |
+
penalty_decay: 0.996
|
| 33 |
+
stop:
|
| 34 |
+
- "\n\n"
|
| 35 |
+
stop_tokens:
|
| 36 |
+
- 0
|
| 37 |
+
- SERVICE_NAME: "rwkv7-g1a-0.4b-20250905-ctx4096"
|
| 38 |
+
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-0.4b-20250905-ctx4096.pth"
|
| 39 |
+
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 40 |
+
REASONING: True
|
| 41 |
+
DEFAULT_SAMPLER:
|
| 42 |
+
max_tokens: 4096
|
| 43 |
+
temperature: 0.1
|
| 44 |
+
top_p: 0.3
|
| 45 |
+
presence_penalty: 0.5
|
| 46 |
+
count_penalty: 0.5
|
| 47 |
+
penalty_decay: 0.996
|
| 48 |
+
stop:
|
| 49 |
+
- "\n\n"
|
| 50 |
+
stop_tokens:
|
| 51 |
+
- 0
|
| 52 |
+
- SERVICE_NAME: "rwkv7-g1a-0.1b-20250728-ctx4096"
|
| 53 |
+
DOWNLOAD_MODEL_FILE_NAME: "rwkv7-g1a-0.1b-20250728-ctx4096.pth"
|
| 54 |
+
DOWNLOAD_MODEL_REPO_ID: "RWKV/rwkv7-g1"
|
| 55 |
+
REASONING: True
|
| 56 |
+
DEFAULT_CHAT: True
|
| 57 |
+
DEFAULT_REASONING: True
|
| 58 |
+
DEFAULT_SAMPLER:
|
| 59 |
+
max_tokens: 4096
|
| 60 |
+
temperature: 0.1
|
| 61 |
+
top_p: 0.3
|
| 62 |
+
presence_penalty: 0.5
|
| 63 |
+
count_penalty: 0.5
|
| 64 |
+
penalty_decay: 0.996
|
| 65 |
+
stop:
|
| 66 |
+
- "\n\n"
|
| 67 |
+
stop_tokens:
|
| 68 |
+
- 0
|