Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

Fix #3275: calibrate a dual config whose gpu is the string 0,1 (from PR #3276; a single card given as a number keeps asking for that card)

MEsfitsddbbda4824f5
diff --git a/setup.py b/setup.py
index bbb3380a1..fec024004 100644
--- a/setup.py
+++ b/setup.py
@@ -1738,7 +1738,7 @@ def recommend_remote_expert_opt(cfg: dict, off: bool = False) -> None:
plain helper path). The engine uses it only with a helper cache; a layer split runs as before. A recommendation:
`off` (setup's --no-remote-expert-opt) or "remote_expert_opt": false in the config keeps it out, and a single-GPU
config is not touched."""
- if not isinstance(cfg.get("gpu"), list) or len(cfg["gpu"]) < 2:
+ if len(gpu_sel(cfg.get("gpu"))) < 2:
return
args = cfg.setdefault("args", [])
if off or cfg.get("remote_expert_opt") is False:
@@ -5949,11 +5949,29 @@ def hip_config_cards(sel) -> list[dict]:
return cards
+def gpu_sel(value) -> list:
+ """A config's "gpu" as a list of card numbers: 0, [0, 2] or the string form "0,2" (which serve splits
+ the same way). [] when it names none. Without this, a string "0,1" reached gpu_info() whole and
+ calibrate died with "there is no GPU 0,1" on a dual-GPU config."""
+ if value is None or value == "":
+ return []
+ items = value if isinstance(value, (list, tuple)) else str(value).split(",")
+ sel = []
+ for x in items:
+ x = str(x).strip()
+ if x == "":
+ continue
+ sel.append(int(x) if x.isdigit() else x)
+ return sel
+
+
def hardware_key(cfg: dict) -> str:
"""What a calibration is valid for: this GPU, CPU and RAM, and the model with its context and images setting
(the context's KV cache and the image encoder take VRAM from the expert cache). #566: a HIP config's cards are
AMD's (hip_config_cards) - nvidia-smi's list named them "?" (or another card with that number) before."""
sel = cfg.get("gpu")
+ if isinstance(sel, str):
+ sel = gpu_sel(sel)
if cfg.get("backend") == "hip":
gl = hip_config_cards(sel)
else:
diff --git a/tools/test_setup_amd.py b/tools/test_setup_amd.py
index 3e1001b8d..6a3f121ae 100644
--- a/tools/test_setup_amd.py
+++ b/tools/test_setup_amd.py
@@ -416,6 +416,22 @@ class CalibrationKey(unittest.TestCase):
"NVIDIA GeForce RTX 5070 + NVIDIA GeForce RTX 5070|24GB|AMD Ryzen 9 7950X|64GB|"
"qwen3.8-flash-next-iq3_xxs|32768|text")
+ def test_nvidia_string_dual_keys_like_the_list(self):
+ """A dual config's string "0,1" keys like [0, 1] (calibrate died with "there is no GPU 0,1")."""
+ def amd(*a, **k):
+ raise AssertionError("amd_gpus called for an NVIDIA config")
+ with mock.patch.object(setup, "amd_gpus", amd), \
+ mock.patch.object(setup, "gpu_info", lambda i=None: {"name": "NVIDIA GeForce RTX 5070", "vram_gb": 11.94}):
+ self.assertEqual(setup.hardware_key(self.cfg("0,1", backend="cuda")),
+ setup.hardware_key(self.cfg([0, 1], backend="cuda")))
+ self.assertEqual(setup.gpu_sel("0,1"), [0, 1])
+ self.assertEqual(setup.gpu_sel("0"), [0])
+ asked = []
+ with mock.patch.object(setup, "amd_gpus", amd), \
+ mock.patch.object(setup, "gpu_info", lambda i=None: asked.append(i) or {"name": "NVIDIA GeForce RTX 5070", "vram_gb": 11.94}):
+ setup.hardware_key(self.cfg(1, backend="cuda")) # a single card given as a number still asks for that card
+ self.assertEqual(asked, [1])
+
def test_setup_reuses_a_saved_calibration_on_linux_hip(self):
cfg = self.cfg(1)
saved = {"settings": {"pcie_frac": 0.0}, "tok_s": 40.1, "date": "2026-10-03"}