fix(hwfit): serve profiles for sub-8192 context models

Allow serve-profile generation for models whose trained context window is below 8192 while preserving the 8K shrink floor for larger models.

services/hwfit/profiles.py

@@ -188,12 +188,18 @@ def compute_serve_profiles(system, model, serve_weights_gb=None, serve_quant=Non
         # Shrink context if even the chosen KV won't fit alongside weights.
         # Start from the smaller of the profile's target and the model's limit.
         cur_ctx = min(ctx, model_ctx_max)
-        while cur_ctx >= 8192:
+        # Floor the context-shrink loop at 8192, but never above the model's own
+        # trained limit. A model with a sub-8192 context (e.g. a 2048-token
+        # SmolLM) starts below 8192, so a hard-coded 8192 guard skipped the loop
+        # entirely and produced NO profile — the serve UI then fell back to
+        # manual flags even though the model fits the GPU trivially.
+        ctx_floor = min(8192, model_ctx_max)
+        while cur_ctx >= ctx_floor:
             kv = _kv_gb(model, cur_ctx, kv_type)
             n_cpu_moe, fits = _cpu_moe_for_budget(model, quant, kv, budget, fixed_gb=serve_weights_gb)
             est = _weights_gb(model, quant, serve_weights_gb) + kv + 0.6
             # If a non-MoE model can't fit even fully offloaded, try less context.
-            if model.get("is_moe") or fits or cur_ctx <= 8192:
+            if model.get("is_moe") or fits or cur_ctx <= ctx_floor:
                 profiles.append({
                     "key": key,
                     "label": label,

tests/test_serve_profiles.py

@@ -81,6 +81,18 @@ def test_context_capped_at_model_limit():
         assert p["ctx"] <= 32768, p
 
 
+def test_small_context_model_still_gets_profiles():
+    """A model whose trained context is below the 8192 shrink floor must still
+    produce serve profiles, capped at its own limit — the loop floor must not
+    exclude it entirely (125 of the catalog models have context_length < 8192)."""
+    small_ctx_model = dict(_DENSE_8B, name="SmolLM-135M", context_length=2048)
+    profs = compute_serve_profiles(_sys(24.0), small_ctx_model)
+    assert profs, "sub-8192-context model produced no profiles"
+    for p in profs:
+        assert p["ctx"] <= 2048, p          # never exceeds the model's trained limit
+        assert p["ctx"] > 0
+
+
 def test_no_gpu_returns_empty():
     """No VRAM detected → no GPU profiles (caller falls back to manual flags)."""
     assert compute_serve_profiles({"backend": "cpu_x86", "gpu_vram_gb": 0}, _QWEN_35B_MOE) == []