konflux: build ramalama images for s390x and ppc64le

.tekton/bats/bats-pull-request.yaml

@@ -26,6 +26,12 @@ spec:
     value: quay.io/redhat-user-workloads/ramalama-tenant/bats:on-pr-{{revision}}
   - name: image-expires-after
     value: 5d
+  - name: build-platforms
+    value:
+    - linux-d160-c8xlarge/amd64
+    - linux-d160-c8xlarge/arm64
+    - linux/ppc64le
+    - linux/s390x
   - name: dockerfile
     value: container-images/bats/Containerfile
   pipelineRef:

.tekton/bats/bats-push.yaml

@@ -23,6 +23,12 @@ spec:
     value: '{{revision}}'
   - name: output-image
     value: quay.io/redhat-user-workloads/ramalama-tenant/bats:{{revision}}
+  - name: build-platforms
+    value:
+    - linux-d160-c8xlarge/amd64
+    - linux-d160-c8xlarge/arm64
+    - linux/ppc64le
+    - linux/s390x
   - name: dockerfile
     value: container-images/bats/Containerfile
   pipelineRef:

.tekton/integration/pipelines/bats-integration.yaml

@@ -17,6 +17,8 @@ spec:
     default:
     - linux-c8xlarge/amd64
     - linux-c8xlarge/arm64
+    - linux/ppc64le
+    - linux/s390x
   - name: commands
     description: Test commands to run
     type: array

.tekton/integration/tasks/test-vm-cmd.yaml

@@ -93,7 +93,7 @@ spec:
       --security-opt label=disable \
       --security-opt unmask=/proc/* \
       --device /dev/net/tun \
-      --device /dev/fuse \
+      -v /tmp \
       ${PODMAN_ENV[*]} \
       $TEST_IMAGE $TEST_CMD
       SCRIPTEOF

.tekton/ramalama/ramalama-pull-request.yaml

@@ -26,6 +26,12 @@ spec:
     value: quay.io/redhat-user-workloads/ramalama-tenant/ramalama:on-pr-{{revision}}
   - name: image-expires-after
     value: 5d
+  - name: build-platforms
+    value:
+    - linux-d160-c8xlarge/amd64
+    - linux-d160-c8xlarge/arm64
+    - linux/ppc64le
+    - linux/s390x
   - name: dockerfile
     value: container-images/ramalama/Containerfile
   - name: test-image

.tekton/ramalama/ramalama-push.yaml

@@ -23,6 +23,12 @@ spec:
     value: '{{revision}}'
   - name: output-image
     value: quay.io/redhat-user-workloads/ramalama-tenant/ramalama:{{revision}}
+  - name: build-platforms
+    value:
+    - linux-d160-c8xlarge/amd64
+    - linux-d160-c8xlarge/arm64
+    - linux/ppc64le
+    - linux/s390x
   - name: dockerfile
     value: container-images/ramalama/Containerfile
   - name: test-image

Makefile

@@ -177,14 +177,15 @@ bats-image:
 	podman inspect $(BATS_IMAGE) &> /dev/null || \
 		podman build -t $(BATS_IMAGE) -f container-images/bats/Containerfile .
 
-bats-in-container: extra-opts = --security-opt unmask=/proc/* --device /dev/net/tun --device /dev/fuse
+bats-in-container: extra-opts = --security-opt unmask=/proc/* --device /dev/net/tun
 
 %-in-container: bats-image
 	podman run -it --rm \
 		--userns=keep-id:size=200000 \
 		--security-opt label=disable \
 		--security-opt=mask=/sys/bus/pci/drivers/i915 \
 		$(extra-opts) \
+		-v /tmp \
 		-v $(CURDIR):/src \
 		$(BATS_IMAGE) make $*
 

container-images/bats/Containerfile

@@ -5,12 +5,13 @@ ENV HOME=/tmp \
 WORKDIR /src
 ENTRYPOINT ["/usr/bin/entrypoint.sh"]
 
-RUN dnf -y install make bats jq iproute podman openssl httpd-tools diffutils \
+RUN dnf -y --setopt=install_weak_deps=false install \
+    make bats jq iproute podman openssl httpd-tools diffutils \
     python3-huggingface-hub \
     $([ $(uname -m) == "x86_64" ] && echo ollama) \
     # for validate and unit-tests
     black codespell shellcheck \
-    python3-flake8 python3-isort python3-pip python3-pytest \
+    python3-flake8 python3-isort python3-pip python3-pytest python3-pyyaml \
     perl-Clone perl-FindBin \
     # for building llama-bench
     git-core cmake gcc-c++ curl-devel && \

container-images/bats/containers.conf

@@ -5,4 +5,3 @@ log_driver="k8s-file"
 
 [engine]
 events_logger="file"
-infra_image="quay.io/libpod/k8s-pause:3.5"

container-images/scripts/build_llama_and_whisper.sh

@@ -66,8 +66,7 @@ dnf_install_rocm() {
   rm_non_ubi_repos
 }
 
-dnf_install_s390() {
-  # I think this was for s390, maybe ppc also
+dnf_install_s390_ppc64le() {
   dnf install -y "openblas-devel"
 }
 
@@ -126,7 +125,7 @@ dnf_install() {
     if [ "$uname_m" = "x86_64" ] || [ "$uname_m" = "aarch64" ]; then
       dnf_install_mesa # on x86_64 and aarch64 we use vulkan via mesa
     else
-      dnf_install_s390
+      dnf_install_s390_ppc64le
     fi
   elif [[ "$containerfile" = rocm* ]]; then
     dnf_install_rocm
@@ -298,9 +297,8 @@ add_common_flags() {
   ramalama)
     if [ "$uname_m" = "x86_64" ] || [ "$uname_m" = "aarch64" ]; then
       common_flags+=("-DGGML_VULKAN=ON")
-    elif [ "$uname_m" = "s390x" ]; then
-      common_flags+=("-DGGML_VXE=ON" "-DGGML_BLAS=ON")
-      common_flags+=("-DGGML_BLAS_VENDOR=OpenBLAS")
+    elif [ "$uname_m" = "s390x" ] || [ "$uname_m" = "ppc64le" ]; then
+      common_flags+=("-DGGML_BLAS=ON" "-DGGML_BLAS_VENDOR=OpenBLAS")
     fi
     ;;
   esac

docs/ramalama-pull.1.md

@@ -20,6 +20,9 @@ Print usage message
 #### **--tls-verify**=*true*
 require HTTPS and verify certificates when contacting OCI registries
 
+#### **--verify**=*true*
+verify the model after pull, disable to allow pulling of models with different endianness
+
 ## SEE ALSO
 **[ramalama(1)](ramalama.1.md)**
 

ramalama/cli.py

@@ -27,6 +27,7 @@
 from ramalama.chat import default_prefix
 from ramalama.common import accel_image, get_accel, perror
 from ramalama.config import CONFIG, coerce_to_bool, load_file_config
+from ramalama.endian import EndianMismatchError
 from ramalama.logger import configure_logger, logger
 from ramalama.model import (
     MODEL_TYPES,
@@ -626,6 +627,12 @@ def pull_parser(subparsers):
         default=True,
         help="require HTTPS and verify certificates when contacting registries",
     )
+    parser.add_argument(
+        "--verify",
+        default=CONFIG.verify,
+        action=CoerceToBool,
+        help="verify the model after pull, disable to allow pulling of models with different endianness",
+    )
     parser.add_argument("MODEL", completer=suppressCompleter)  # positional argument
     parser.set_defaults(func=pull_cli)
 
@@ -1427,6 +1434,8 @@ def eprint(e, exit_code):
         eprint(e, errno.ENOSYS)
     except subprocess.CalledProcessError as e:
         eprint(e, e.returncode)
+    except EndianMismatchError:
+        sys.exit(1)
     except KeyboardInterrupt:
         sys.exit(0)
     except IOError as e:

ramalama/config.py

@@ -106,6 +106,7 @@ class BaseConfig:
     threads: int = -1
     transport: str = "ollama"
     user: UserConfig = field(default_factory=UserConfig)
+    verify: bool = True
 
     def __post_init__(self):
         self.container = coerce_to_bool(self.container) if self.container is not None else self.engine is not None
@@ -189,7 +190,7 @@ def load_env_config(env: Mapping[str, str] | None = None) -> dict[str, Any]:
     if 'images' in config:
         config['images'] = json.loads(config['images'])
 
-    for key in ['ocr', 'keep_groups', 'container']:
+    for key in ['ocr', 'keep_groups', 'container', 'verify']:
         if key in config:
             config[key] = coerce_to_bool(config[key])
 

ramalama/hf_style_repo_base.py

@@ -14,6 +14,7 @@
     perror,
     run_cmd,
 )
+from ramalama.endian import EndianMismatchError
 from ramalama.logger import logger
 from ramalama.model import Model
 from ramalama.model_store.snapshot_file import SnapshotFile, SnapshotFileType
@@ -264,8 +265,11 @@ def pull(self, args):
             repo = self.create_repository(name, organization, tag)
             snapshot_hash = repo.model_hash
             files = repo.get_file_list(cached_files)
-            self.model_store.new_snapshot(tag, snapshot_hash, files)
+            self.model_store.new_snapshot(tag, snapshot_hash, files, verify=getattr(args, "verify", True))
 
+        except EndianMismatchError:
+            # No use pulling again
+            raise
         except Exception as e:
             if not available(self.get_cli_command()):
                 perror(f"URL pull failed and {self.get_cli_command()} not available")
@@ -278,7 +282,7 @@ def pull(self, args):
                 run_cmd(conman_args)
 
                 snapshot_hash, files = self._collect_cli_files(tempdir)
-                self.model_store.new_snapshot(tag, snapshot_hash, files)
+                self.model_store.new_snapshot(tag, snapshot_hash, files, verify=getattr(args, "verify", True))
 
     def exec(self, cmd_args, args):
         try:

ramalama/model_store/store.py

@@ -309,7 +309,7 @@ def verify_snapshot(self, model_tag: str):
         self._verify_endianness(model_tag)
         self._store.verify_snapshot()
 
-    def new_snapshot(self, model_tag: str, snapshot_hash: str, snapshot_files: list[SnapshotFile]):
+    def new_snapshot(self, model_tag: str, snapshot_hash: str, snapshot_files: list[SnapshotFile], verify: bool = True):
         snapshot_hash = sanitize_filename(snapshot_hash)
 
         try:
@@ -327,7 +327,8 @@ def new_snapshot(self, model_tag: str, snapshot_hash: str, snapshot_files: list[
             raise ex
 
         try:
-            self.verify_snapshot(model_tag)
+            if verify:
+                self.verify_snapshot(model_tag)
         except EndianMismatchError as ex:
             perror(f"Verification of snapshot failed: {ex}")
             perror("Removing snapshot...")

ramalama/ollama.py

@@ -188,7 +188,7 @@ def pull(self, args) -> None:
             self.print_pull_message(f"ollama://{organization}/{name}:{tag}")
 
         model_hash = ollama_repo.get_model_hash(manifest)
-        self.model_store.new_snapshot(tag, model_hash, files)
+        self.model_store.new_snapshot(tag, model_hash, files, verify=getattr(args, "verify", True))
 
         # If a model has been downloaded via ollama cli, only create symlink in the snapshots directory
         if is_model_in_ollama_cache:

ramalama/url.py

@@ -104,7 +104,7 @@ def _assemble_split_file_list(self, snapshot_hash: str) -> list[SnapshotFile]:
 
         return files
 
-    def pull(self, _) -> None:
+    def pull(self, args) -> None:
         name, tag, _ = self.extract_model_identifiers()
         _, _, all_files = self.model_store.get_cached_files(tag)
         if all_files:
@@ -122,12 +122,12 @@ def pull(self, _) -> None:
                     required=True,
                 )
             )
-            self.model_store.new_snapshot(tag, snapshot_hash, files)
+            self.model_store.new_snapshot(tag, snapshot_hash, files, verify=getattr(args, "verify", True))
             return
 
         if is_split_file_model(self.model):
             files = self._assemble_split_file_list(snapshot_hash)
-            self.model_store.new_snapshot(tag, snapshot_hash, files)
+            self.model_store.new_snapshot(tag, snapshot_hash, files, verify=getattr(args, "verify", True))
             return
 
         files.append(
@@ -141,5 +141,5 @@ def pull(self, _) -> None:
                 required=True,
             )
         )
-        self.model_store.new_snapshot(tag, snapshot_hash, files)
+        self.model_store.new_snapshot(tag, snapshot_hash, files, verify=getattr(args, "verify", True))
         return

shortnames/shortnames.conf

@@ -24,6 +24,8 @@
   "granite:2b" = "ollama://granite3.1-dense:2b"
   "granite:7b" = "huggingface://instructlab/granite-7b-lab-GGUF/granite-7b-lab-Q4_K_M.gguf"
   "granite:8b" = "ollama://granite3.1-dense:8b"
+  "granite-be-3.0:1b" = "hf://taronaeo/Granite-3.0-1B-A400M-Instruct-BE-GGUF/granite-3.0-1b-a400m-instruct-be.Q2_K.gguf"
+  "granite-be-3.3:2b" = "hf://taronaeo/Granite-3.3-2B-Instruct-BE-GGUF/granite-3.3-2b-instruct-be.Q4_K_M.gguf"
   "hermes" = "huggingface://NousResearch/Hermes-2-Pro-Mistral-7B-GGUF/Hermes-2-Pro-Mistral-7B.Q4_K_M.gguf"
   "ibm/granite" = "ollama://granite3.1-dense:8b"
   "ibm/granite:2b" = "ollama://granite3.1-dense:2b"

test/system/002-bench.bats

@@ -16,7 +16,10 @@ function setup() {
 # bats test_tags=distro-integration
 @test "ramalama bench" {
     skip_if_no_llama_bench
-    run_ramalama bench -t 2 smollm:135m
+    if is_s390x; then
+        local RAMALAMA_TIMEOUT=1200
+    fi
+    run_ramalama bench -t 2 $(test_model smollm:135m)
     is "$output" ".*model.*size.*" "model and size in output"
 }
 

test/system/030-run.bats

@@ -125,19 +125,20 @@ EOF
     is "$output" ".*-e HSA_OVERRIDE_GFX_VERSION=0.0.0" "ensure HSA_OVERRIDE_GFX_VERSION is set from environment"
 }
 
-@test "ramalama run smollm with prompt" {
-    run_ramalama run --temp 0 ${MODEL} "What is the first line of the declaration of independence?"
+@test "ramalama run with prompt" {
+    run_ramalama run --temp 0 $(test_model ${MODEL} granite-be-3.3:2b) "What is the first line of the declaration of independence?"
 }
 
 @test "ramalama run --keepalive" {
     # timeout within 1 second and generate a 124 error code.
-    run_ramalama 0 --debug run --keepalive 1s tiny
+    run_ramalama 0 --debug run --keepalive 1s $(test_model tiny)
 }
 
 @test "ramalama run --image bogus" {
     skip_if_nocontainer
     skip_if_darwin
     skip_if_docker
+    run_ramalama pull tiny
     run_ramalama 22 run --image bogus --pull=never tiny
     is "$output" ".*Error: bogus: image not known"
     run_ramalama 125 run --image bogus1 --rag quay.io/ramalama/rag --pull=never tiny

test/system/040-serve.bats

@@ -126,7 +126,9 @@ verify_begin=".*run --rm"
     run_ramalama -q --dryrun serve smollm
     is "$output" ".*ai.ramalama.model=ollama://library/smollm:latest" "smollm should be expanded to fullname"
 
-    model=ollama://smollm:135m
+    model=$(test_model smollm:135m)
+    run_ramalama info
+    full_model=$(jq -r --arg model $model -r '.Shortnames.Names[$model]' <<<"$output")
     container1=c_$(safename)
     container2=c_$(safename)
 
@@ -139,7 +141,7 @@ verify_begin=".*run --rm"
     port=${output: -8:4}
 
     run_ramalama chat --ls --url http://127.0.0.1:${port}/v1
-    is "$output" "smollm:135m" "list of models available correct"
+    is "$output" ${full_model#*://} "list of models available correct"
 
     run_ramalama containers --noheading
     is "$output" ".*${container1}" "list correct for container1"
@@ -159,7 +161,7 @@ verify_begin=".*run --rm"
 @test "ramalama --detach serve multiple" {
     skip_if_nocontainer
 
-    model=ollama://smollm:135m
+    model=$(test_model ollama://smollm:135m)
     container=c_$(safename)
     port1=8100
     port2=8200
@@ -445,6 +447,8 @@ verify_begin=".*run --rm"
 @test "ramalama serve --api llama-stack" {
     skip_if_docker
     skip_if_nocontainer
+    skip_if_ppc64le
+    skip_if_s390x
     model=tiny
     name=c_$(safename)
     run_ramalama pull ${model}

test/system/050-pull.bats

@@ -167,6 +167,24 @@ load setup_suite
     run_ramalama rm oci://quay.io/mmortari/gguf-py-example:v1
 }
 
+@test "ramalama pull little-endian" {
+    if ! is_bigendian; then
+        skip "Testing pulls of opposite-endian models"
+    fi
+    run_ramalama rm --ignore tiny
+    run_ramalama 1 pull --verify=on tiny
+    is "$output" ".*Endian mismatch of host (BIG) and model (LITTLE).*" "detected little-endian model"
+}
+
+@test "ramalama pull big-endian" {
+    if is_bigendian; then
+        skip "Testing pulls of opposite-endian models"
+    fi
+    run_ramalama rm --ignore granite-be-3.0:1b
+    run_ramalama 1 pull --verify=on granite-be-3.0:1b
+    is "$output" ".*Endian mismatch of host (LITTLE) and model (BIG).*" "detected big-endian model"
+}
+
 @test "ramalama URL" {
       model=$RAMALAMA_TMPDIR/mymodel.gguf
       touch $model