vimlm 0.0.4__tar.gz → 0.0.5__tar.gz

{vimlm-0.0.4/vimlm.egg-info → vimlm-0.0.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: vimlm
-Version: 0.0.4
+Version: 0.0.5
 Summary: VimLM - LLM-powered Vim assistant
 Home-page: https://github.com/JosefAlbers/vimlm
 Author: Josef Albers
@@ -8,7 +8,8 @@ Author-email: albersj66@gmail.com
 Requires-Python: >=3.12.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: nanollama==0.0.3b0
+Requires-Dist: nanollama==0.0.5b0
+Requires-Dist: mlx_lm_utils==0.0.1a0
 Requires-Dist: watchfiles==1.0.4
 Dynamic: author
 Dynamic: author-email
@@ -19,7 +20,7 @@ Dynamic: requires-dist
 Dynamic: requires-python
 Dynamic: summary
 
-# VimLM - Vim Language Model Assistant for privacy-conscious developers
+# VimLM - Local LLM-Powered Coding Assistant for Vim
 
 ![vimlm](https://github.com/user-attachments/assets/67a97048-48df-40c1-9109-53c759e85d96)
 
@@ -27,11 +28,20 @@ An LLM-powered coding companion for Vim, inspired by GitHub Copilot/Cursor. Inte
 
 ## Features
 
-- **Real-Time Interaction with local LLMs**: Runs **fully offline** with local models (default: uncensored Llama-3.2-3B).
-- **Integration with Vim's Native Workflow**: Simple Vim keybindings for quick access and split-window interface for non-intrusive responses.
-- **Context-Awareness Beyond Single Files**: Inline support for external documents and project files for richer, more informed responses.
-- **Conversational AI Assistance**: Goes beyond simple code suggestions-explains reasoning, provides alternative solutions, and allows interactive follow-ups. 
-- **Versatile Use Cases**: Not just for coding-use it for documentation lookup, general Q&A, or even casual (uncensored) conversations.
+- **Model Agnostic** - Use any MLX-compatible model via config file
+- **Vim-Native UX** - Ctrl-l/Ctrl-r keybindings and split-window responses
+- **Deep Context** - Understands code context from:
+    - Current file
+    - Visual selections
+    - Referenced files (`!@#$` syntax)
+    - Project directory structure
+- **Conversational Coding** - Iterative refinement with follow-up queries
+- **Air-Gapped Security** - 100% offline - no APIs, no tracking, no data leaks
+
+## Requirements
+
+- Apple M-series chip (M1/M2/M3/M4)
+- Python 3.12.8
 
 ## Installation
 
@@ -39,15 +49,9 @@ An LLM-powered coding companion for Vim, inspired by GitHub Copilot/Cursor. Inte
 pip install vimlm
 ```
 
-## Usage
+## Quick Start
 
-1. Start Vim with VimLM:
-
-```zsh
-vimlm
-```
-
-or
+1. Launch with default model (DeepSeek-R1-Distill-Qwen-7B-4bit):
 
 ```zsh
 vimlm your_file.js
@@ -58,7 +62,7 @@ vimlm your_file.js
     - Example prompt: "Regex for removing html tags in item.content"
 
 3. **From Visual Mode**:
-    - Select text → `Ctrl-l`: Send selection + file context
+    - Select code → `Ctrl-l`: Send selection + file context
     - Example prompt: "Convert this to async/await syntax"
 
 4. **Add Context**: Use `!@#$` to include additional files/folders:
@@ -71,6 +75,28 @@ vimlm your_file.js
     - `Ctrl-r`: Continue thread
     - Example follow-up: "In Manifest V3"
 
+## Advanced Configuration
+
+### Custom Model Setup
+
+1. **Browse models**: [MLX Community Models on Hugging Face](https://huggingface.co/mlx-community)
+
+2. **Edit config file**:
+
+```json
+{
+ "LLM_MODEL": "/path/to/your/mlx_model"
+}
+```
+
+3. **Save to**:
+
+```
+~/vimlm/cfg.json
+```
+
+4. **Restart VimLM**
+
 ## Key Bindings
 
 | Binding    | Mode          | Action                                 |

{vimlm-0.0.4 → vimlm-0.0.5}/README.md

@@ -1,4 +1,4 @@
-# VimLM - Vim Language Model Assistant for privacy-conscious developers
+# VimLM - Local LLM-Powered Coding Assistant for Vim
 
 ![vimlm](https://github.com/user-attachments/assets/67a97048-48df-40c1-9109-53c759e85d96)
 
@@ -6,11 +6,20 @@ An LLM-powered coding companion for Vim, inspired by GitHub Copilot/Cursor. Inte
 
 ## Features
 
-- **Real-Time Interaction with local LLMs**: Runs **fully offline** with local models (default: uncensored Llama-3.2-3B).
-- **Integration with Vim's Native Workflow**: Simple Vim keybindings for quick access and split-window interface for non-intrusive responses.
-- **Context-Awareness Beyond Single Files**: Inline support for external documents and project files for richer, more informed responses.
-- **Conversational AI Assistance**: Goes beyond simple code suggestions-explains reasoning, provides alternative solutions, and allows interactive follow-ups. 
-- **Versatile Use Cases**: Not just for coding-use it for documentation lookup, general Q&A, or even casual (uncensored) conversations.
+- **Model Agnostic** - Use any MLX-compatible model via config file
+- **Vim-Native UX** - Ctrl-l/Ctrl-r keybindings and split-window responses
+- **Deep Context** - Understands code context from:
+    - Current file
+    - Visual selections
+    - Referenced files (`!@#$` syntax)
+    - Project directory structure
+- **Conversational Coding** - Iterative refinement with follow-up queries
+- **Air-Gapped Security** - 100% offline - no APIs, no tracking, no data leaks
+
+## Requirements
+
+- Apple M-series chip (M1/M2/M3/M4)
+- Python 3.12.8
 
 ## Installation
 
@@ -18,15 +27,9 @@ An LLM-powered coding companion for Vim, inspired by GitHub Copilot/Cursor. Inte
 pip install vimlm
 ```
 
-## Usage
+## Quick Start
 
-1. Start Vim with VimLM:
-
-```zsh
-vimlm
-```
-
-or
+1. Launch with default model (DeepSeek-R1-Distill-Qwen-7B-4bit):
 
 ```zsh
 vimlm your_file.js
@@ -37,7 +40,7 @@ vimlm your_file.js
     - Example prompt: "Regex for removing html tags in item.content"
 
 3. **From Visual Mode**:
-    - Select text → `Ctrl-l`: Send selection + file context
+    - Select code → `Ctrl-l`: Send selection + file context
     - Example prompt: "Convert this to async/await syntax"
 
 4. **Add Context**: Use `!@#$` to include additional files/folders:
@@ -50,6 +53,28 @@ vimlm your_file.js
     - `Ctrl-r`: Continue thread
     - Example follow-up: "In Manifest V3"
 
+## Advanced Configuration
+
+### Custom Model Setup
+
+1. **Browse models**: [MLX Community Models on Hugging Face](https://huggingface.co/mlx-community)
+
+2. **Edit config file**:
+
+```json
+{
+ "LLM_MODEL": "/path/to/your/mlx_model"
+}
+```
+
+3. **Save to**:
+
+```
+~/vimlm/cfg.json
+```
+
+4. **Restart VimLM**
+
 ## Key Bindings
 
 | Binding    | Mode          | Action                                 |

{vimlm-0.0.4 → vimlm-0.0.5}/setup.py

@@ -5,7 +5,7 @@ with open("requirements.txt") as f:
 
 setup(
     name="vimlm",
-    version="0.0.4",
+    version="0.0.5",
     author="Josef Albers",
     author_email="albersj66@gmail.com",
     readme='README.md',

{vimlm-0.0.4 → vimlm-0.0.5/vimlm.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: vimlm
-Version: 0.0.4
+Version: 0.0.5
 Summary: VimLM - LLM-powered Vim assistant
 Home-page: https://github.com/JosefAlbers/vimlm
 Author: Josef Albers
@@ -8,7 +8,8 @@ Author-email: albersj66@gmail.com
 Requires-Python: >=3.12.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: nanollama==0.0.3b0
+Requires-Dist: nanollama==0.0.5b0
+Requires-Dist: mlx_lm_utils==0.0.1a0
 Requires-Dist: watchfiles==1.0.4
 Dynamic: author
 Dynamic: author-email
@@ -19,7 +20,7 @@ Dynamic: requires-dist
 Dynamic: requires-python
 Dynamic: summary
 
-# VimLM - Vim Language Model Assistant for privacy-conscious developers
+# VimLM - Local LLM-Powered Coding Assistant for Vim
 
 ![vimlm](https://github.com/user-attachments/assets/67a97048-48df-40c1-9109-53c759e85d96)
 
@@ -27,11 +28,20 @@ An LLM-powered coding companion for Vim, inspired by GitHub Copilot/Cursor. Inte
 
 ## Features
 
-- **Real-Time Interaction with local LLMs**: Runs **fully offline** with local models (default: uncensored Llama-3.2-3B).
-- **Integration with Vim's Native Workflow**: Simple Vim keybindings for quick access and split-window interface for non-intrusive responses.
-- **Context-Awareness Beyond Single Files**: Inline support for external documents and project files for richer, more informed responses.
-- **Conversational AI Assistance**: Goes beyond simple code suggestions-explains reasoning, provides alternative solutions, and allows interactive follow-ups. 
-- **Versatile Use Cases**: Not just for coding-use it for documentation lookup, general Q&A, or even casual (uncensored) conversations.
+- **Model Agnostic** - Use any MLX-compatible model via config file
+- **Vim-Native UX** - Ctrl-l/Ctrl-r keybindings and split-window responses
+- **Deep Context** - Understands code context from:
+    - Current file
+    - Visual selections
+    - Referenced files (`!@#$` syntax)
+    - Project directory structure
+- **Conversational Coding** - Iterative refinement with follow-up queries
+- **Air-Gapped Security** - 100% offline - no APIs, no tracking, no data leaks
+
+## Requirements
+
+- Apple M-series chip (M1/M2/M3/M4)
+- Python 3.12.8
 
 ## Installation
 
@@ -39,15 +49,9 @@ An LLM-powered coding companion for Vim, inspired by GitHub Copilot/Cursor. Inte
 pip install vimlm
 ```
 
-## Usage
+## Quick Start
 
-1. Start Vim with VimLM:
-
-```zsh
-vimlm
-```
-
-or
+1. Launch with default model (DeepSeek-R1-Distill-Qwen-7B-4bit):
 
 ```zsh
 vimlm your_file.js
@@ -58,7 +62,7 @@ vimlm your_file.js
     - Example prompt: "Regex for removing html tags in item.content"
 
 3. **From Visual Mode**:
-    - Select text → `Ctrl-l`: Send selection + file context
+    - Select code → `Ctrl-l`: Send selection + file context
     - Example prompt: "Convert this to async/await syntax"
 
 4. **Add Context**: Use `!@#$` to include additional files/folders:
@@ -71,6 +75,28 @@ vimlm your_file.js
     - `Ctrl-r`: Continue thread
     - Example follow-up: "In Manifest V3"
 
+## Advanced Configuration
+
+### Custom Model Setup
+
+1. **Browse models**: [MLX Community Models on Hugging Face](https://huggingface.co/mlx-community)
+
+2. **Edit config file**:
+
+```json
+{
+ "LLM_MODEL": "/path/to/your/mlx_model"
+}
+```
+
+3. **Save to**:
+
+```
+~/vimlm/cfg.json
+```
+
+4. **Restart VimLM**
+
 ## Key Bindings
 
 | Binding    | Mode          | Action                                 |

vimlm-0.0.5/vimlm.egg-info/requires.txt

@@ -0,0 +1,3 @@
+nanollama==0.0.5b0
+mlx_lm_utils==0.0.1a0
+watchfiles==1.0.4

{vimlm-0.0.4 → vimlm-0.0.5}/vimlm.py

@@ -17,7 +17,6 @@ import subprocess
 import json
 import os
 from watchfiles import awatch
-from nanollama32 import Chat
 import shutil
 import time
 from itertools import accumulate
@@ -27,6 +26,7 @@ from pathlib import Path
 from string import Template
 
 DEBUG = True
+LLM_MODEL = "mlx-community/DeepSeek-R1-Distill-Qwen-7B-4bit"
 NUM_TOKEN = 2000
 SEP_CMD = '!@#$'
 VIMLM_DIR = os.path.expanduser("~/vimlm")
@@ -41,20 +41,6 @@ LOG_PATH = os.path.join(VIMLM_DIR, LOG_FILE)
 LTM_PATH = os.path.join(VIMLM_DIR, LTM_FILE)
 OUT_PATH = os.path.join(WATCH_DIR, OUT_FILE) 
 
-if os.path.exists(WATCH_DIR):
-    shutil.rmtree(WATCH_DIR)
-os.makedirs(WATCH_DIR)
-
-try:
-    with open(CFG_PATH, "r") as f:
-        cfg = cfg.load(f)
-    DEBUG = config.get("DEBUG", DEBUG)
-    NUM_TOKEN = config.get("NUM_TOKEN", NUM_TOKEN)
-    SEP_CMD = config.get("SEP_CMD", SEP_CMD)
-except:
-    with open(CFG_PATH, 'w') as f:
-        json.dump(dict(DEBUG=DEBUG, NUM_TOKEN=NUM_TOKEN, SEP_CMD=SEP_CMD), f, indent=2)
-
 def toout(s, key='tovim'):
     with open(OUT_PATH, 'w', encoding='utf-8') as f:
         f.write(s)
@@ -63,18 +49,40 @@ def toout(s, key='tovim'):
 def tolog(log, key='debug'):
     if not DEBUG and key == 'debug':
         return
-    if os.path.exists(LOG_PATH):
+    try:
         with open(LOG_PATH, "r", encoding="utf-8") as log_f:
             logs = json.load(log_f)
-    else:
+    except:
         logs = []
     logs.append(dict(key=key, log=log, timestamp=time.ctime()))
     with open(LOG_PATH, "w", encoding="utf-8") as log_f:
         json.dump(logs, log_f, indent=2)
 
+if os.path.exists(WATCH_DIR):
+    shutil.rmtree(WATCH_DIR)
+os.makedirs(WATCH_DIR)
+
+try:
+    with open(CFG_PATH, "r") as f:
+        config = json.load(f)
+    DEBUG = config.get("DEBUG", DEBUG)
+    LLM_MODEL = config.get("LLM_MODEL", LLM_MODEL)
+    NUM_TOKEN = config.get("NUM_TOKEN", NUM_TOKEN)
+    SEP_CMD = config.get("SEP_CMD", SEP_CMD)
+except Exception as e:
+    tolog(str(e))
+    with open(CFG_PATH, 'w') as f:
+        json.dump(dict(DEBUG=DEBUG, LLM_MODEL=LLM_MODEL, NUM_TOKEN=NUM_TOKEN, SEP_CMD=SEP_CMD), f, indent=2)
+
 toout('Loading LLM...')
-chat = Chat(variant='uncn_llama_32_3b_it')
-toout('LLM is ready')
+if LLM_MODEL is None:
+    from nanollama32 import Chat
+    chat = Chat(variant='uncn_llama_32_3b_it')
+    toout('LLM is ready')
+else:
+    from mlx_lm_utils import Chat
+    chat = Chat(model_path=LLM_MODEL)
+    toout(f'{LLM_MODEL.split('/')[-1]} is ready')
 
 def is_binary(file_path):
     try:
@@ -142,9 +150,6 @@ def retrieve(src_path, max_len=2000, get_len=len):
             continue
     return result
 
-def get_ntok(s):
-    return len(chat.tokenizer.encode(s)[0])
-
 def ingest(src):
     def load_cache(cache_path=LTM_PATH):
         if os.path.exists(cache_path):
@@ -161,7 +166,7 @@ def ingest(src):
     toout('Ingesting...')
     format_ingest = '{volat}{incoming}\n\n---\n\nPlease provide a succint bullet point summary for above:'
     format_volat = 'Here is a summary of part 1 of **{k}**:\n\n---\n\n{newsum}\n\n---\n\nHere is the next part:\n\n---\n\n'
-    dict_doc = retrieve(src, get_len=get_ntok)
+    dict_doc = retrieve(src, get_len=chat.get_ntok)
     dict_sum = {}
     cache = load_cache()
     for k, v in dict_doc.items():
@@ -178,14 +183,14 @@ def ingest(src):
             accum = ''
             for s in list_str:
                 chat.reset()
-                newsum = chat(format_ingest.format(volat=volat, incoming=s.strip()), max_new=max_new_sum, verbose=False, stream=None)[0][:-10].strip()
+                newsum = chat(format_ingest.format(volat=volat, incoming=s.strip()), max_new=max_new_sum, verbose=False, stream=None)['text']
                 accum += newsum + ' ...\n'
                 volat = format_volat.format(k=k, newsum=newsum)
         else:
             accum = list_str[0]
         chat.reset()
         toout('')
-        chat_summary = chat(format_ingest.format(volat=f'**{k}**:\n', incoming=accum), max_new=int(NUM_TOKEN/4), verbose=False, stream=OUT_PATH)[0][:-10].strip()
+        chat_summary = chat(format_ingest.format(volat=f'**{k}**:\n', incoming=accum), max_new=int(NUM_TOKEN/4), verbose=False, stream=OUT_PATH)['text']
         dict_sum[k] = dict(timestamp=v_stamp, summary=chat_summary)
     dump_cache(dict_sum)
     result = ''
@@ -253,9 +258,9 @@ async def process_files(data):
     prompt = str_template.format(**data)
     tolog(prompt, 'tollm')
     toout('')
-    response = chat(prompt, max_new=NUM_TOKEN - get_ntok(prompt), verbose=False, stream=OUT_PATH)
-    toout(response[0][:-10].strip())
-    tolog(response[-1], 'tps')
+    response = chat(prompt, max_new=NUM_TOKEN - chat.get_ntok(prompt), verbose=False, stream=OUT_PATH)
+    toout(response['text'])
+    tolog(response['benchmark'])
 
 VIMLMSCRIPT = Template(r"""
 let s:watched_dir = expand('$WATCH_DIR')

vimlm-0.0.4/vimlm.egg-info/requires.txt

@@ -1,2 +0,0 @@
-nanollama==0.0.3b0
-watchfiles==1.0.4