work in progress to support GCP H100's

src/packages/frontend/compute/action.tsx

@@ -1,6 +1,5 @@
 import { Alert, Button, Modal, Popconfirm, Popover, Spin } from "antd";
 import { useEffect, useState } from "react";
-
 import { redux, useStore } from "@cocalc/frontend/app-framework";
 import { A, CopyToClipBoard, Icon } from "@cocalc/frontend/components";
 import ShowError from "@cocalc/frontend/components/error";
@@ -13,6 +12,7 @@ import {
   ACTION_INFO,
   STATE_INFO,
   getTargetState,
+  getArchitecture,
 } from "@cocalc/util/db-schema/compute-servers";
 import { computeServerAction, getApiKey } from "./api";
 import costPerHour from "./cost";
@@ -62,7 +62,7 @@ export default function getActions({
       if (configuration.cloud != "google-cloud") {
         continue;
       }
-      if (configuration.machineType.startsWith("t2a-")) {
+      if (getArchitecture(configuration) == "arm64") {
         // TODO: suspend/resume breaks the clock badly on ARM64, and I haven't
         // figured out a workaround, so don't support it for now.  I guess this
         // is a GCP bug.

src/packages/frontend/compute/google-cloud-config.tsx

@@ -7,7 +7,10 @@ import type {
 } from "@cocalc/util/db-schema/compute-servers";
 import { reloadImages, useImages, useGoogleImages } from "./images-hook";
 import { GOOGLE_CLOUD_DEFAULTS } from "@cocalc/util/db-schema/compute-servers";
-import { getMinDiskSizeGb } from "@cocalc/util/db-schema/compute-servers";
+import {
+  getMinDiskSizeGb,
+  getArchitecture,
+} from "@cocalc/util/db-schema/compute-servers";
 import {
   Alert,
   Button,
@@ -301,9 +304,7 @@ export default function GoogleCloudConfiguration({
             !!(configuration.acceleratorType && configuration.acceleratorCount)
           }
           googleImages={googleImages}
-          arch={
-            configuration.machineType?.startsWith("t2a-") ? "arm64" : "x86_64"
-          }
+          arch={getArchitecture(configuration)}
         />
       ),
     },
@@ -937,25 +938,23 @@ function Zone({ priceData, setConfig, configuration, disabled }) {
 
 function MachineType({ priceData, setConfig, configuration, disabled, state }) {
   const [archType, setArchType] = useState<"x86_64" | "arm64">(
-    configuration.machineType?.startsWith("t2a-") ? "arm64" : "x86_64",
+    getArchitecture(configuration),
   );
   const [sortByPrice, setSortByPrice] = useState<boolean>(true);
   const [newMachineType, setNewMachineType] = useState<string>(
     configuration.machineType ?? "",
   );
   useEffect(() => {
     setNewMachineType(configuration.machineType);
-    setArchType(
-      configuration.machineType?.startsWith("t2a-") ? "arm64" : "x86_64",
-    );
+    setArchType(getArchitecture(configuration));
   }, [configuration.machineType]);
   useEffect(() => {
-    if (archType == "arm64" && !configuration.machineType.startsWith("t2a-")) {
+    if (archType == "arm64" && getArchitecture(configuration) != "arm64") {
       setNewMachineType("t2a-standard-4");
       setConfig({ machineType: "t2a-standard-4" });
       return;
     }
-    if (archType == "x86_64" && configuration.machineType.startsWith("t2a-")) {
+    if (archType == "x86_64" && getArchitecture(configuration) == "arm64") {
       setNewMachineType("t2d-standard-4");
       setConfig({ machineType: "t2d-standard-4" });
       return;
@@ -967,13 +966,13 @@ function MachineType({ priceData, setConfig, configuration, disabled, state }) {
     .filter((machineType) => {
       const { acceleratorType } = configuration;
       if (!acceleratorType) {
-        if (machineType.startsWith("g2-") || machineType.startsWith("a2-")) {
+        if (machineType.startsWith("g") || machineType.startsWith("a")) {
           return false;
         }
-        if (archType == "arm64" && !machineType.startsWith("t2a-")) {
+        if (archType == "arm64" && getArchitecture(configuration) != "arm64") {
           return false;
         }
-        if (archType == "x86_64" && machineType.startsWith("t2a-")) {
+        if (archType == "x86_64" && getArchitecture(configuration) == "arm64") {
           return false;
         }
       } else {
@@ -1052,8 +1051,8 @@ function MachineType({ priceData, setConfig, configuration, disabled, state }) {
             (state ?? "deprovisioned") != "deprovisioned"
               ? "Can only be changed when machine is deprovisioned"
               : archType == "x86_64"
-              ? "Intel or AMD X86_64 architecture machines"
-              : "ARM64 architecture machines"
+                ? "Intel or AMD X86_64 architecture machines"
+                : "ARM64 architecture machines"
           }
         >
           <Radio.Group
@@ -1161,6 +1160,9 @@ const ACCELERATOR_TYPES = [
   "nvidia-l4",
   "nvidia-tesla-a100",
   "nvidia-a100-80gb",
+  "nvidia-h100-80gb",
+  "nvidia-h200-141gb",
+  // these are too hard to properly keep software image for:
   // "nvidia-tesla-v100",
   //"nvidia-tesla-p100",
   //"nvidia-tesla-p4",
@@ -1188,12 +1190,16 @@ function GPU({
     <div style={{ color: "#666", marginBottom: "5px" }}>
       <b>
         <Icon style={{ float: "right", fontSize: "50px" }} name="gpu" />
-        <Icon name="cube" /> NVIDIA GPU:{" "}
-        <A href="https://www.nvidia.com/en-us/data-center/a100/">A100</A>,{" "}
-        <A href="https://www.nvidia.com/en-us/data-center/l4/">L4</A>,{" "}
-        <A href="https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/solutions/resources/documents1/Datasheet_NVIDIA_T4_Virtualization.pdf">
-          T4
-        </A>
+        <Icon name="cube" /> NVIDIA GPU{" "}
+        <div style={{ float: "right" }}>
+          <A href="https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/solutions/resources/documents1/Datasheet_NVIDIA_T4_Virtualization.pdf">
+            T4
+          </A>
+          , <A href="https://www.nvidia.com/en-us/data-center/l4/">L4</A>,{" "}
+          <A href="https://www.nvidia.com/en-us/data-center/a100/">A100</A>,{" "}
+          <A href="https://www.nvidia.com/en-us/data-center/h100/">H100</A>,{" "}
+          <A href="https://www.nvidia.com/en-us/data-center/h200/">H200</A>
+        </div>
       </b>
     </div>
   );
@@ -1377,23 +1383,16 @@ function ensureConsistentConfiguration(
 ) {
   const newConfiguration = { ...configuration, ...changes };
   const newChanges = { ...changes };
-
   ensureConsistentImage(newConfiguration, newChanges, IMAGES);
-
   ensureConsistentAccelerator(priceData, newConfiguration, newChanges);
-
   ensureConsistentNvidiaL4andA100(priceData, newConfiguration, newChanges);
-
+  ensureConsistentZoneWithRegion(priceData, newConfiguration, newChanges);
   ensureConsistentRegionAndZoneWithMachineType(
     priceData,
     newConfiguration,
     newChanges,
   );
-
-  ensureConsistentZoneWithRegion(priceData, newConfiguration, newChanges);
-
   ensureSufficientDiskSize(newConfiguration, newChanges, IMAGES);
-
   ensureConsistentDiskType(priceData, newConfiguration, newChanges);
 
   return newChanges;
@@ -1453,13 +1452,15 @@ function ensureConsistentZoneWithRegion(priceData, configuration, changes) {
     // currently changing region, so set a zone that matches the region
     for (const zone in priceData.zones) {
       if (zone.startsWith(configuration.region)) {
-        changes["zone"] = zone;
+        configuration["zone"] = changes["zone"] = zone;
         break;
       }
     }
   } else {
     // probably changing the zone, so set the region from the zone
-    changes["region"] = zoneToRegion(configuration.zone);
+    configuration["region"] = changes["region"] = zoneToRegion(
+      configuration.zone,
+    );
   }
 }
 
@@ -1562,13 +1563,17 @@ function ensureZoneIsConsistentWithGPU(priceData, configuration, changes) {
   }
 }
 
-// The Nvidia L4 and A100 are a little different
+// The Nvidia L4 and A100 are a little different, etc.
 function ensureConsistentNvidiaL4andA100(priceData, configuration, changes) {
   const { machineType, acceleratorType } = configuration;
 
   // L4 or A100 GPU machine type, but switching to no GPU, so we have
   // to change the machine type
-  if (machineType.startsWith("g2-") || machineType.startsWith("a2-")) {
+  if (
+    machineType.startsWith("g2-") ||
+    machineType.startsWith("a2-") ||
+    machineType.startsWith("a3-")
+  ) {
     if (!acceleratorType) {
       // Easy case -- the user is explicitly changing the GPU from being set
       // to NOT be set, and the GPU is L4 or A100.  In this case,
@@ -1579,6 +1584,8 @@ function ensureConsistentNvidiaL4andA100(priceData, configuration, changes) {
     }
   }
   if (
+    acceleratorType != "nvidia-h200-141gb" &&
+    acceleratorType != "nvidia-h100-80gb" &&
     acceleratorType != "nvidia-tesla-a100" &&
     acceleratorType != "nvidia-a100-80gb" &&
     acceleratorType != "nvidia-l4"
@@ -1602,6 +1609,21 @@ function ensureConsistentNvidiaL4andA100(priceData, configuration, changes) {
       priceData.accelerators[acceleratorType]?.machineType[
         configuration.acceleratorCount
       ];
+
+    if (machineTypes == null) {
+      // maybe 1 gpu isn't allowed, e.g., with H200
+      const machineType = priceData.accelerators[acceleratorType]?.machineType;
+      if (machineType != null) {
+        for (const count in machineType) {
+          configuration.acceleratorCount = changes.acceleratorCount =
+            parseInt(count);
+          machineTypes =
+            priceData.accelerators[acceleratorType]?.machineType[
+              configuration.acceleratorCount
+            ];
+        }
+      }
+    }
   }
   if (machineTypes == null) {
     throw Error("bug -- this can't happen");
@@ -1626,7 +1648,7 @@ function ensureConsistentRegionAndZoneWithMachineType(
   const machineType = configuration["machineType"];
   if (priceData.machineTypes[machineType] == null) {
     console.warn(
-      `BUG -- This should never happen: unknonwn machineType = '${machineType}'`,
+      `BUG -- This should never happen: unknown machineType = '${machineType}'`,
     );
     // invalid machineType
     if (configuration.acceleratorType) {

src/packages/server/compute/cloud/google-cloud/images.ts

@@ -27,6 +27,7 @@ import type {
   GoogleCloudImage,
   GoogleCloudImages,
 } from "@cocalc/util/db-schema/compute-servers";
+import { getArchitecture as getArchitecture0 } from "@cocalc/util/db-schema/compute-servers";
 import { makeValidGoogleName } from "@cocalc/util/db-schema/compute-servers";
 import { cmp } from "@cocalc/util/misc";
 import { getGoogleCloudImagePrefix } from "./index";
@@ -158,7 +159,7 @@ export async function deleteImage(name: string) {
 }
 
 export function getArchitecture(machineType: string): Architecture {
-  return machineType.startsWith("t2a-") ? "arm64" : "x86_64";
+  return getArchitecture0({ machineType, cloud: "google-cloud" } as any);
 }
 
 export async function getSourceImage({

src/packages/util/compute/cloud/clouds.ts

@@ -42,7 +42,7 @@ export const GOOGLE_CLOUD_DEFAULTS = {
     machineType: "n2d-highmem-2",
     spot: GCLOUD_SPOT_DEFAULT,
     diskSizeGb: 10,
-    diskType: "pd-balanced",
+    diskType: "pd-standard",
     externalIp: true,
     excludeFromSync: DEFAULT_EXCLUDE_FROM_SYNC,
   },

src/packages/util/db-schema/compute-servers.ts

@@ -251,11 +251,30 @@ for (const state of ORDERED_STATES) {
   n += 1;
 }
 
+export function getArchitecture(configuration: Configuration): Architecture {
+  if (configuration.cloud == "onprem") {
+    return configuration.arch ?? "x86_64";
+  }
+  if (configuration.cloud != "google-cloud") {
+    // no ARM outside of GCP right now
+    return "x86_64";
+  }
+  const { machineType } = configuration;
+  const v = machineType.split("-");
+  if (v[0].endsWith("a")) {
+    // The known machines with are are: t2a-, c4a-
+    // Everything else ends with a number or d.
+    // Hopefully this pattern persists.
+    return "arm64";
+  }
+  return "x86_64";
+}
+
 function supportsSuspend(configuration: Configuration) {
   if (configuration.cloud != "google-cloud") {
     return false;
   }
-  if (configuration.machineType.startsWith("t2a-")) {
+  if (getArchitecture(configuration) != "x86_64") {
     // TODO: suspend/resume breaks the clock badly on ARM64, and I haven't
     // figured out a workaround, so don't support it for now.  I guess this
     // is a GCP bug.
@@ -699,6 +718,8 @@ interface FluidStackConfiguration extends BaseConfiguration {
   os: string;
 }
 export type GoogleCloudAcceleratorType =
+  | "nvidia-h200-141gb"
+  | "nvidia-h100-80gb"
   | "nvidia-a100-80gb"
   | "nvidia-tesla-a100"
   | "nvidia-l4"
@@ -708,6 +729,8 @@ export type GoogleCloudAcceleratorType =
   | "nvidia-tesla-p100";
 
 export const GOOGLE_CLOUD_ACCELERATOR_TYPES: GoogleCloudAcceleratorType[] = [
+  "nvidia-h200-141gb",
+  "nvidia-h100-80gb",
   "nvidia-a100-80gb",
   "nvidia-tesla-a100",
   "nvidia-l4",