From f3d5544849fa2e61966dd50cc9f2fa1546b419c0 Mon Sep 17 00:00:00 2001
From: sd109 <sdavidson327@gmail.com>
Date: Thu, 31 Oct 2024 17:44:09 +0000
Subject: [PATCH] Remove top_k from vision model UI options

---
 charts/azimuth-image-analysis/azimuth-ui.schema.yaml | 4 +++-
 charts/azimuth-image-analysis/ci/ui-only-values.yaml | 1 -
 charts/azimuth-image-analysis/values.schema.json     | 8 +-------
 web-apps/image-analysis/defaults.yml                 | 4 +++-
 4 files changed, 7 insertions(+), 10 deletions(-)

diff --git a/charts/azimuth-image-analysis/azimuth-ui.schema.yaml b/charts/azimuth-image-analysis/azimuth-ui.schema.yaml
index 885ca8e..a960081 100644
--- a/charts/azimuth-image-analysis/azimuth-ui.schema.yaml
+++ b/charts/azimuth-image-analysis/azimuth-ui.schema.yaml
@@ -27,5 +27,7 @@ sortOrder:
   - /azimuth-llm/ui/appSettings/llm_params/frequency_penalty
   - /azimuth-llm/ui/appSettings/llm_params/presence_penalty
   - /azimuth-llm/ui/appSettings/llm_params/top_p
-  - /azimuth-llm/ui/appSettings/llm_params/top_k
+  # vLLM responds with HTTP 400 BadRequest when top_k is
+  # passed to a vision model (but ollama accepts it)
+  # - /azimuth-llm/ui/appSettings/llm_params/top_k
   - /azimuth-llm/api/modelMaxContextLength
diff --git a/charts/azimuth-image-analysis/ci/ui-only-values.yaml b/charts/azimuth-image-analysis/ci/ui-only-values.yaml
index b66347d..96f716d 100644
--- a/charts/azimuth-image-analysis/ci/ui-only-values.yaml
+++ b/charts/azimuth-image-analysis/ci/ui-only-values.yaml
@@ -11,6 +11,5 @@ azimuth-llm:
         max_tokens: 101
         temperature: 0.1
         top_p: 0.15
-        top_k: 1
         presence_penalty: 0.9
         frequency_penalty: 1
diff --git a/charts/azimuth-image-analysis/values.schema.json b/charts/azimuth-image-analysis/values.schema.json
index 7f747e4..c8be1ac 100644
--- a/charts/azimuth-image-analysis/values.schema.json
+++ b/charts/azimuth-image-analysis/values.schema.json
@@ -58,6 +58,7 @@
                                     "description": "Model name supplied to the OpenAI client in frontend web app. Should match huggingface.model above."
                                 },
                                 "llm_params": {
+                                    "$comment": "top_k parameter causes vLLM to error for most (all?) vision models so is excluded here",
                                     "type": "object",
                                     "properties": {
                                         "max_tokens": {
@@ -82,13 +83,6 @@
                                             "exclusiveMinimum": 0,
                                             "maximum": 1
                                         },
-                                        "top_k": {
-                                            "type": "integer",
-                                            "title": "LLM Top K",
-                                            "description": "The [top k](https://docs.vllm.ai/en/stable/dev/sampling_params.html) value to use when generating LLM responses (must be an integer).",
-                                            "default": -1,
-                                            "minimum": -1
-                                        },
                                         "presence_penalty": {
                                             "type": "number",
                                             "title": "LLM Presence Penalty",
diff --git a/web-apps/image-analysis/defaults.yml b/web-apps/image-analysis/defaults.yml
index e6f2791..21d233c 100644
--- a/web-apps/image-analysis/defaults.yml
+++ b/web-apps/image-analysis/defaults.yml
@@ -18,7 +18,9 @@ llm_params:
   max_tokens:
   temperature:
   top_p:
-  top_k:
+  # vLLM rejects requests with top_k parameter for
+  # most (all?) vision models so can't use it here
+  # top_k:
   frequency_penalty:
   presence_penalty: