Add H200 and B200 support for Crusoe provider (#216)

Author: peterschmidt85

src/gpuhunt/_internal/constraints.py

@@ -177,6 +177,7 @@ def is_nvidia_superchip(gpu_name: str) -> bool:
     NvidiaGPUInfo(name="H200", memory=141, compute_capability=(9, 0)),
     NvidiaGPUInfo(name="H200NVL", memory=141, compute_capability=(9, 0)),
     NvidiaGPUInfo(name="B200", memory=180, compute_capability=(10, 0)),
+    NvidiaGPUInfo(name="GB200", memory=186, compute_capability=(10, 0)),
     NvidiaGPUInfo(name="L4", memory=24, compute_capability=(8, 9)),
     NvidiaGPUInfo(name="L40", memory=48, compute_capability=(8, 9)),
     NvidiaGPUInfo(name="L40S", memory=48, compute_capability=(8, 9)),
@@ -253,6 +254,12 @@ def is_nvidia_superchip(gpu_name: str) -> bool:
         architecture=AMDArchitecture.CDNA3,
         device_ids=(0x74A5,),
     ),
+    AMDGPUInfo(
+        name="MI355X",
+        memory=288,
+        architecture=AMDArchitecture.CDNA4,
+        device_ids=(0x75A3,),
+    ),
 ]
 
 KNOWN_TPUS: list[TPUInfo] = [TPUInfo(name=version, memory=0) for version in _TPU_VERSIONS]

src/gpuhunt/_internal/models.py

@@ -32,6 +32,7 @@ class AMDArchitecture(enum.Enum):
     CDNA = "CDNA"
     CDNA2 = "CDNA2"
     CDNA3 = "CDNA3"
+    CDNA4 = "CDNA4"
 
     @classmethod
     def cast(cls, value: Union["AMDArchitecture", str]) -> "AMDArchitecture":

src/gpuhunt/providers/crusoe.py

@@ -10,7 +10,12 @@
 
 import requests
 
-from gpuhunt._internal.models import AcceleratorVendor, QueryFilter, RawCatalogItem
+from gpuhunt._internal.models import (
+    AcceleratorVendor,
+    CPUArchitecture,
+    QueryFilter,
+    RawCatalogItem,
+)
 from gpuhunt.providers import AbstractProvider
 
 logger = logging.getLogger(__name__)
@@ -26,15 +31,13 @@
     "A100-PCIe-80GB": ("A100", AcceleratorVendor.NVIDIA, 80),
     "A100-SXM-80GB": ("A100", AcceleratorVendor.NVIDIA, 80),
     "H100-SXM-80GB": ("H100", AcceleratorVendor.NVIDIA, 80),
+    "H200-SXM-141GB": ("H200", AcceleratorVendor.NVIDIA, 141),
+    "B200-SXM-180GB": ("B200", AcceleratorVendor.NVIDIA, 180),
+    "GB200-NVL-186GB": ("GB200", AcceleratorVendor.NVIDIA, 186),
     "L40S-48GB": ("L40S", AcceleratorVendor.NVIDIA, 48),
     "A40-PCIe-48GB": ("A40", AcceleratorVendor.NVIDIA, 48),
     "MI300X-192GB": ("MI300X", AcceleratorVendor.AMD, 192),
-    # TODO: The following GPUs are listed on https://crusoe.ai/cloud/pricing but not yet
-    # returned by the instance types API. Add them once Crusoe exposes them:
-    #   - H200 141GB ($4.29/GPU-hr on-demand, spot: contact sales)
-    #   - GB200 186GB (contact sales)
-    #   - B200 180GB (contact sales)
-    #   - MI355X 288GB ($3.45 listed but not confirmed; also missing from KNOWN_AMD_GPUS)
+    "MI355X_288GB": ("MI355X", AcceleratorVendor.AMD, 288),
 }
 
 # Per-GPU-hour pricing from https://crusoe.ai/cloud/pricing
@@ -44,6 +47,10 @@
     "A100-PCIe-80GB": (1.65, 1.20),
     "A100-SXM-80GB": (1.95, 1.30),
     "H100-SXM-80GB": (3.90, 1.60),
+    "H200-SXM-141GB": (4.29, None),
+    # TODO: B200 estimated from B200/H100 ratio on other providers; update once Crusoe publishes rates.
+    # GB200 and MI355X pricing not known yet; update once Crusoe publishes rates.
+    "B200-SXM-180GB": (7.25, None),
     "L40S-48GB": (1.00, 0.50),
     "A40-PCIe-48GB": (0.90, 0.40),
     "MI300X-192GB": (3.45, 0.95),
@@ -137,6 +144,13 @@ def _request(self, method: str, path: str, params: Optional[dict] = None) -> req
         return requests.request(method, url, headers=headers, params=params, timeout=TIMEOUT)
 
 
+def _get_cpu_arch(spec: dict) -> str:
+    cpu_type = spec.get("cpu_type", "")
+    if cpu_type == "arm64":
+        return CPUArchitecture.ARM.value
+    return CPUArchitecture.X86.value
+
+
 def _get_available_type_locations(capacities: list[dict]) -> dict[str, list[str]]:
     best_qty: dict[tuple[str, str], int] = defaultdict(int)
     for cap in capacities:
@@ -191,6 +205,7 @@ def _make_gpu_items(
         gpu_memory=gpu_memory,
         spot=None,
         disk_size=float(spec["disk_gb"]) if spec.get("disk_gb") else None,
+        cpu_arch=_get_cpu_arch(spec),
         # disk_gb: ephemeral NVMe size in GB (0 = no ephemeral disk).
         # Used by dstack to decide whether to create a persistent data disk.
         provider_data={"disk_gb": spec.get("disk_gb", 0)},
@@ -231,6 +246,7 @@ def _make_cpu_items(product_name: str, spec: dict, locations: list[str]) -> list
         gpu_memory=None,
         spot=False,
         disk_size=float(spec["disk_gb"]) if spec.get("disk_gb") else None,
+        cpu_arch=_get_cpu_arch(spec),
         provider_data={"disk_gb": spec.get("disk_gb", 0)},
     )