fix(llm): enable flash-attn on qwen-classifier, real VRAM cause found

Batch/ubatch reduction barely moved measured VRAM (~768MB free, same as
before) — wrong lever. llama-server runs with --flash-attn on; this
service didn't. Without it, the unfused attention compute buffer at
65536 ctx is far larger than flash-attn's fused workspace, which is
what the naive weights+KV estimate missed. Matches llama-server's flag.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

docker-compose.yml

@@ -68,15 +68,17 @@ services:
   #
   # VRAM: weights+KV math (~4.9GiB) predicted comfortable headroom in the
   # ~6.1GiB free on the R9700, but measured live it actually used ~5.85GiB —
-  # left only ~700MB free, too tight (compute-buffer/graph overhead the
-  # naive weights+KV estimate didn't account for). --batch-size/--ubatch-size
-  # dropped well below llama-server's defaults (2048/512) to shrink that
-  # buffer — this classifier processes one request at a time, no batching
-  # throughput to lose. --n-gpu-layers 999 (full offload) kept since the
-  # ~1.5GB freed here should cover it; re-verify with rocm-smi after deploy
-  # before trusting this number again. GPU_MAX_HW_QUEUES=1 carried over from
-  # llama-server's comment above — same ROCm/ROCm#5706 clock-pinning bug
-  # applies now that two HIP contexts (this + llama-server) share the card.
+  # left only ~700MB free, too tight. Dropping --batch-size/--ubatch-size
+  # barely moved it (~768MB free) — wrong lever. Actual cause: llama-server
+  # runs with --flash-attn on but this service was missing it — without
+  # flash attention the unfused attention compute buffer at 65536 ctx is
+  # much larger (roughly O(n^2) intermediate buffers vs flash-attn's fused,
+  # near-linear workspace), dwarfing the naive weights+KV estimate. Added
+  # --flash-attn on to match llama-server; re-verify with rocm-smi after
+  # deploy before trusting any of these numbers again. GPU_MAX_HW_QUEUES=1
+  # carried over from llama-server's comment above — same ROCm/ROCm#5706
+  # clock-pinning bug applies now that two HIP contexts (this +
+  # llama-server) share the card.
   qwen-classifier:
     image: ghcr.io/ggml-org/llama.cpp:server-rocm
     container_name: qwen-classifier
@@ -102,6 +104,7 @@ services:
       --parallel 1
       --batch-size 512
       --ubatch-size 128
+      --flash-attn on
       --cache-type-k q4_0
       --cache-type-v q4_0
       --jinja