fix(hwfit): detect unified-memory NVIDIA (Grace Blackwell GB10 / DGX Spark) instead of 'No GPU' (#1340) (#1372)

_detect_nvidia parsed nvidia-smi --query-gpu=memory.total,name and did
float(memory.total) per row, dropping the row on ValueError. Grace Blackwell
GB10 (DGX Spark, sm_121) reports memory.total as '[N/A]'/'Not Supported'
because the GPU shares the system LPDDR pool rather than carrying discrete VRAM
— so the only GPU row was dropped and a real GB10 (even with vLLM running on it)
was reported as 'No GPU', breaking Cookbook recommendations and model switching.

Keep a named device whose memory.total is non-numeric: when there are no
discrete-VRAM rows but such unified devices exist, report a unified-memory CUDA
GPU backed by the system RAM pool (has_gpu, name, backend=cuda, count,
unified_memory=True) — mirroring how Apple Silicon and AMD APUs are already
handled. Discrete GPUs are unchanged, and a box with a real discrete GPU keeps
the discrete path.

Adds tests/test_hwfit_unified_nvidia.py with a GB10 nvidia-smi fixture: the
device is detected (not dropped), surfaces through detect_system with
unified_memory propagated, discrete GPUs stay non-unified, and a discrete GPU
takes precedence over an N/A-memory row.

Co-authored-by: NubsCarson <nubs@nubs.site>

services/hwfit/hardware.py

@@ -105,6 +105,8 @@ def _detect_nvidia():
         return None
 
     gpus = []
+    # Devices nvidia-smi lists with a real name but a non-numeric memory.total.
+    unified = []
     # nvidia-smi lists GPUs in index order (0,1,2,...), so the row position is
     # the CUDA device index we'd pass to CUDA_VISIBLE_DEVICES.
     for idx, line in enumerate(out.strip().split("\n")):
@@ -114,9 +116,32 @@ def _detect_nvidia():
                 vram_mb = float(parts[0])
                 gpus.append({"index": idx, "name": parts[1], "vram_gb": vram_mb / 1024.0})
             except ValueError:
+                # Grace Blackwell GB10 / DGX Spark and other unified-memory
+                # NVIDIA parts report memory.total as "[N/A]"/"Not Supported"
+                # because the GPU shares the system LPDDR pool instead of
+                # carrying discrete VRAM. Don't drop the device — remember it so
+                # we report a unified-memory GPU below rather than "No GPU" (#1340).
+                if parts[1]:
+                    unified.append({"index": idx, "name": parts[1]})
                 continue
 
     if not gpus:
+        if unified:
+            # Unified-memory CUDA box: report the GPU backed by system RAM so the
+            # Cookbook recommends models and serving works. The pool is shared
+            # (not per-GPU discrete VRAM), so report the RAM total once.
+            ram_gb = round(_get_ram_gb(), 1)
+            gpus = [{"index": g["index"], "name": g["name"], "vram_gb": ram_gb} for g in unified]
+            return {
+                "gpu_name": gpus[0]["name"],
+                "gpu_vram_gb": ram_gb,
+                "gpu_count": len(gpus),
+                "gpus": gpus,
+                "gpu_groups": _group_gpus(gpus),
+                "homogeneous": True,
+                "backend": "cuda",
+                "unified_memory": True,
+            }
         return None
     total_vram = sum(g["vram_gb"] for g in gpus)
     groups = _group_gpus(gpus)