copy

aider/website/_data/quant.yml

@@ -24,7 +24,7 @@
 
 - dirname: 2024-11-22-18-56-13--ollama-qwen2.5-coder:32b-instruct-fp16
   test_cases: 132
-  model: ollama/qwen2.5-coder:32b-instruct-fp16 (64k context)
+  model: Ollama fp16
   edit_format: diff
   commit_hash: f06452c-dirty, 6a0a97c-dirty, 4e9ae16-dirty, 5506d0f-dirty
   pass_rate_1: 58.3
@@ -70,7 +70,7 @@
 
 - dirname: 2024-11-22-17-53-35--qwen25-coder-32b-Instruct-4bit
   test_cases: 133
-  model: mlx-community/Qwen2.5-Coder-32B-Instruct-4bit
+  model: mlx-community 4bit
   edit_format: diff
   commit_hash: a16dcab-dirty
   pass_rate_1: 60.2
@@ -93,7 +93,7 @@
 
 - dirname: 2024-11-23-15-07-20--qwen25-coder-32b-Instruct-8bit
   test_cases: 133
-  model: mlx-community/Qwen2.5-Coder-32B-Instruct-8bit
+  model: mlx-community 8bit
   edit_format: diff
   commit_hash: a16dcab-dirty
   pass_rate_1: 59.4
@@ -137,26 +137,25 @@
   seconds_per_case: 40.7
   total_cost: 0.1497
 
-- dirname: 2024-11-21-23-33-47--ollama-qwen25-coder
+- dirname: 2024-11-23-21-08-53--ollama-qwen2.5-coder:32b-instruct-q4_K_M-8kctx
   test_cases: 133
-  model: Ollama Q4_K_M
+  model: Ollama q4_K_M
   edit_format: diff
-  commit_hash: 488c88d-dirty
-  pass_rate_1: 44.4
-  pass_rate_2: 53.4
-  percent_cases_well_formed: 44.4
-  error_outputs: 231
-  num_malformed_responses: 183
-  num_with_malformed_responses: 74
-  user_asks: 79
+  commit_hash: baa1335-dirty, e63df83-dirty, ff8c1aa-dirty
+  pass_rate_1: 54.9
+  pass_rate_2: 66.9
+  percent_cases_well_formed: 94.0
+  error_outputs: 21
+  num_malformed_responses: 21
+  num_with_malformed_responses: 8
+  user_asks: 5
   lazy_comments: 0
-  syntax_errors: 2
+  syntax_errors: 0
   indentation_errors: 0
   exhausted_context_windows: 0
-  test_timeouts: 2
+  test_timeouts: 3
   command: aider --model ollama/qwen2.5-coder:32b-instruct-q4_K_M
-  date: 2024-11-21
+  date: 2024-11-23
   versions: 0.64.2.dev
-  seconds_per_case: 86.7
-  total_cost: 0.0000
-
+  seconds_per_case: 35.7
+  total_cost: 0.0000

aider/website/_posts/2024-11-21-quantization.md

@@ -18,7 +18,7 @@ can strongly impact code editing skill.
 Heavily quantized models are often used by cloud API providers
 and local model servers like Ollama or MLX.
 
-<canvas id="quantChart" width="800" height="600" style="margin: 20px 0"></canvas>
+<canvas id="quantChart" width="800" height="500" style="margin: 20px 0"></canvas>
 <script src="https://cdn.jsdelivr.net/npm/chart.js"></script>
 <script>
 {% include quant-chart.js %}
@@ -29,16 +29,16 @@ served both locally and from cloud providers.
 
 - The [HuggingFace BF16 weights](https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct) served via [glhf.chat](https://glhf.chat).
 - Hyperbolic labs API for [qwen2-5-coder-32b-instruct](https://app.hyperbolic.xyz/models/qwen2-5-coder-32b-instruct), which is using BF16. This result is probably within the expected variance of the HF result.
-- A [4bit quant for mlx](https://t.co/cwX3DYX35D).
+- [4bit and 8bit quants for mlx](https://t.co/cwX3DYX35D).
 - The results from [OpenRouter's mix of providers](https://openrouter.ai/qwen/qwen-2.5-coder-32b-instruct/providers) which serve the model with different levels of quantization.
-- Ollama locally serving [qwen2.5-coder:32b-instruct-q4_K_M)](https://ollama.com/library/qwen2.5-coder:32b-instruct-q4_K_M), which has `Q4_K_M` quantization, with Ollama's default 2k context window.
+- Ollama locally serving different quantizations from the [Ollama model library](https://ollama.com/library/qwen2.5-coder:32b-instruct-q4_K_M).
 
 The best version of the model rivals GPT-4o, while the worst performer
-is more like GPT-3.5 Turbo level.
+is more like GPT-4 level.
 
 {: .note }
 This article is being updated as additional benchmark runs complete.
-The original version included incorrect Ollama models.
+
 
 <input type="text" id="quantSearchInput" placeholder="Search..." style="width: 100%; max-width: 800px; margin: 10px auto; padding: 8px; display: block; border: 1px solid #ddd; border-radius: 4px;">
 
@@ -100,27 +100,30 @@ document.getElementById('quantSearchInput').addEventListener('keyup', function()
 });
 </script>
 
-## Setting the context window size
+## Setting Ollama's context window size
 
 [Ollama uses a 2k context window by default](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-specify-the-context-window-size),
 which is very small for working with aider.
 
+All of the Ollama results above were collected with at least an 8k context window, which
+is large enough to attempt all the coding problems in the benchmark.
+
 You can set the Ollama server's context window with a 
 [`.aider.model.settings.yml` file](https://aider.chat/docs/config/adv-model-settings.html#model-settings)
 like this:
 
 ```
 - name: aider/extra_params
   extra_params:
-    num_ctx: 65536
+    num_ctx: 8192
 ```
 
 That uses the special model name `aider/extra_params` to set it for *all* models. You should probably use a specific model name like:
 
 ```
 - name: ollama/qwen2.5-coder:32b-instruct-fp16
   extra_params:
-    num_ctx: 65536
+    num_ctx: 8192
 ```
 
 ## Choosing providers with OpenRouter
@@ -130,3 +133,8 @@ OpenRouter allows you to ignore specific providers in your
 This can be effective to exclude highly quantized or otherwise
 undesirable providers.
 
+
+{: .note }
+Earlier versions of this article included incorrect Ollama models,
+and also included some Ollama results with the too small default 2k
+context window.