From 90e08f05027916a72a4530547cfe51342b528cf3 Mon Sep 17 00:00:00 2001
From: Vuong Hoang <vh@phasefinal.com>
Date: Thu, 4 Jun 2026 00:29:05 -0700
Subject: [PATCH] fix(vllm-phi4): Ollama-matching chat template to recover R15
 baseline

vLLM's official Phi-4 tokenizer template emits <|end|> after the system turn;
Ollama's does not. That single boundary token regressed brokkr's R15 P02
admission eval (type macro-F1 -33pp) vs the Ollama-measured canonical, while
valid_format held at 1.0. Operator chose to make vLLM match Ollama's leaner
scaffold globally (baseline == production). Adds conf/phi4-chat-template.jinja
(drops the system <|end|>) + mounts it + --chat-template on vllm-phi4. Applied
prompt verified via tokenize/detokenize; brokkr re-smokes probe_vllm.yaml.
---
 stacks/vllm/compose.yaml                  |  5 +++++
 stacks/vllm/conf/phi4-chat-template.jinja | 24 +++++++++++++++++++++++
 2 files changed, 29 insertions(+)
 create mode 100644 stacks/vllm/conf/phi4-chat-template.jinja

diff --git a/stacks/vllm/compose.yaml b/stacks/vllm/compose.yaml
index 1a0334d..28ec332 100644
--- a/stacks/vllm/compose.yaml
+++ b/stacks/vllm/compose.yaml
@@ -210,6 +210,7 @@ services:
       - "${PHI4_PORT}:8000"
     volumes:
       - /tank/aimodels/huggingface:/hfcache
+      - /opt/docker/conf/vllm/phi4-chat-template.jinja:/config/phi4-chat-template.jinja:ro
     environment:
       - HF_HOME=/hfcache
       - HF_HUB_CACHE=/hfcache/hub
@@ -237,6 +238,10 @@ services:
       # FP8 KV cache — halves KV memory at 128K ctx on Ada (cc 8.9); near-lossless.
       - --kv-cache-dtype
       - ${PHI4_KV_CACHE_DTYPE}
+      # Ollama-matching scaffold (drops the system-turn <|end|>) so vLLM reproduces
+      # brokkr's R15 canonical baseline. See conf/phi4-chat-template.jinja.
+      - --chat-template
+      - /config/phi4-chat-template.jinja
     deploy:
       resources:
         reservations:
diff --git a/stacks/vllm/conf/phi4-chat-template.jinja b/stacks/vllm/conf/phi4-chat-template.jinja
new file mode 100644
index 0000000..31b930b
--- /dev/null
+++ b/stacks/vllm/conf/phi4-chat-template.jinja
@@ -0,0 +1,24 @@
+{#-
+  Phi-4-mini chat template — OLLAMA-MATCHING variant.
+
+  Why this exists: the official HF tokenizer template emits <|end|> after the
+  SYSTEM turn; Ollama's phi4 template does NOT (its system-block <|end|> is only
+  in the tools branch). That single boundary token regressed brokkr's R15 P02
+  admission eval on vLLM vs the Ollama-measured canonical (type macro-F1 -33pp)
+  while valid_format held at 1.0. Operator chose (2026-06-04) to make vLLM match
+  Ollama's leaner scaffold globally so baseline == production.
+
+  Renders (system + user, add_generation_prompt):
+    <|system|>{sys}<|user|>{usr}<|end|><|assistant|>
+  i.e. NO <|end|> after the system turn (the only delta from official).
+-#}
+{%- for message in messages -%}
+{%- if message['role'] == 'system' -%}
+{{- '<|system|>' + message['content'] -}}
+{%- else -%}
+{{- '<|' + message['role'] + '|>' + message['content'] + '<|end|>' -}}
+{%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+{{- '<|assistant|>' -}}
+{%- endif -%}