diff --git a/gpu-validation/defaults/main.yaml b/gpu-validation/defaults/main.yaml
index 529ad5e..2e6f724 100644
--- a/gpu-validation/defaults/main.yaml
+++ b/gpu-validation/defaults/main.yaml
@@ -121,3 +121,52 @@ gpu_validation_reboot_force: false
 gpu_validation_reboot_timeout: 3600
 # [int] Post-reboot delay in seconds
 gpu_validation_reboot_post_delay: 60
+
+# ── NVSwitch validation defaults (nvswitch_validation.yaml playbook) ──────────
+# gpu_validation_nvswitch_validation_fabric_mode is intentionally NOT defaulted here.
+# tasks_from in include_role is rendered in the play context before role defaults
+# are loaded, so a role default cannot serve as a reliable selector.
+# Callers MUST supply gpu_validation_nvswitch_validation_fabric_mode via extra-vars (-e),
+# inventory/group_vars, or an AnsibleTest CR ansibleVarFiles entry.
+# Valid values: 0 (PCI grouping), 1 (Shared NVSwitch), 2 (vGPU mdev).
+
+# [list] Systemd services expected active per mode.
+# Override if your installation uses different service names.
+gpu_validation_nvswitch_validation_mode0_fm_services:
+  - nvidia-fabricmanager
+  - nvlsm
+gpu_validation_nvswitch_validation_mode1_fm_services:
+  - nvidia-fabricmanager
+  - nvlsm
+  - fm-partition-restore
+gpu_validation_nvswitch_validation_mode2_fm_services:
+  - nvidia-fabricmanager
+  - nvlsm
+  - nvidia-vgpu-mgr
+
+# [int] Mode 2 only: expected number of mdev instances (from mdevctl list).
+# Set to 0 to skip the assertion. Typically equals gpu_validation_nvswitch_validation_expected_gpu_count.
+gpu_validation_nvswitch_validation_mode2_expected_mdev_count: 0
+
+# [string] GPU PCI vendor ID
+gpu_validation_nvswitch_validation_gpu_vendor_id: "10de"
+# [string] GPU PCI product ID — HGX B300 GB110 SXM6 AC
+gpu_validation_nvswitch_validation_gpu_product_id: "3182"
+# [int] Total number of GPUs expected on the compute host
+gpu_validation_nvswitch_validation_expected_gpu_count: 8
+# [bool] Check that all GPUs reach the expected Fabric State
+gpu_validation_nvswitch_validation_fabric_state_check: true
+# [string] Expected nvidia-smi Fabric State value
+gpu_validation_nvswitch_validation_fabric_state_expected: "Completed"
+# [int] Expected number of activated NVSwitch partitions; set after first fmpm -l run
+gpu_validation_nvswitch_validation_expected_partition_count: 0
+# [int] Number of GPUs per passthrough VM flavor
+gpu_validation_nvswitch_validation_gpus_per_vm: 1
+# [string] Nova resource class for the GPU passthrough flavor
+gpu_validation_nvswitch_validation_passthrough_resource_class: "CUSTOM_HGX-B300-GPU"
+# [string] Expected libvirt device type for passthrough GPUs in virsh dumpxml
+gpu_validation_nvswitch_validation_expected_device_type: "hostdev"
+# [bool] Check that FABRIC_MODE in fabricmanager.cfg matches gpu_validation_nvswitch_validation_fabric_mode
+gpu_validation_nvswitch_validation_check_fabric_mode_cfg: false
+# [string] Path to the FM configuration file
+gpu_validation_nvswitch_validation_fabric_mode_cfg_file: /usr/share/nvidia/nvswitch/fabricmanager.cfg
diff --git a/gpu-validation/tasks/nvswitch.yaml b/gpu-validation/tasks/nvswitch.yaml
new file mode 100644
index 0000000..7e1baa6
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch.yaml
@@ -0,0 +1,83 @@
+---
+# host-side checks: FM services, fabric state, partition activation.
+
+- name: Collect FM service status
+  become: true
+  ansible.builtin.systemd:
+    name: "{{ item }}"
+  loop: "{{ gpu_validation_nvswitch_validation_fm_services }}"
+  register: _fm_service_status
+  failed_when: false
+
+- name: Set FM service status facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fm_service_results: "{{ _fm_service_status.results
+      | map(attribute='item') | zip(_fm_service_status.results
+         | map(attribute='status') | map(attribute='ActiveState'))
+      | list }}"
+
+- name: Show FM service states
+  ansible.builtin.debug:
+    msg: "{{ item[0] }}: {{ item[1] }}"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Collect Fabric State for all GPUs
+  become: true
+  ansible.builtin.shell: |
+    set -o pipefail
+    nvidia-smi -q | grep -A2 'Fabric' | grep 'State' | awk '{print $NF}'
+  args:
+    executable: /bin/bash
+  register: _fabric_states
+  changed_when: false
+  failed_when: false
+
+- name: Show fabric states
+  ansible.builtin.debug:
+    msg: "GPU Fabric States: {{ _fabric_states.stdout_lines }}"
+
+- name: Set fabric state facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fabric_states: "{{ _fabric_states.stdout_lines | map('trim') | list }}"
+    gpu_validation_nvswitch_fabric_completed_count: >-
+      {{ _fabric_states.stdout_lines
+         | map('trim')
+         | select('equalto', gpu_validation_nvswitch_validation_fabric_state_expected)
+         | list | length }}
+
+- name: Collect fmpm partition list
+  become: true
+  ansible.builtin.command:
+    cmd: fmpm -l
+  register: _fmpm_list
+  changed_when: false
+  failed_when: false
+
+- name: Show fmpm partition list
+  ansible.builtin.debug:
+    msg: "{{ _fmpm_list.stdout | default('fmpm not available or no output') }}"
+
+- name: Set partition facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fmpm_output: "{{ _fmpm_list.stdout | default('') }}"
+    gpu_validation_nvswitch_activated_partition_count: >-
+      {{ (_fmpm_list.stdout | default(''))
+         .splitlines()
+         | select('search', '\\bactivated\\b')
+         | list | length }}
+
+- name: Collect NVLink topology matrix
+  become: true
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _nvlink_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show NVLink topology matrix
+  ansible.builtin.debug:
+    msg: "{{ _nvlink_topo.stdout | default('nvidia-smi topo not available') }}"
+
+- name: Set topology fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_topo_output: "{{ _nvlink_topo.stdout | default('') }}"
diff --git a/gpu-validation/tasks/nvswitch_assertions.yaml b/gpu-validation/tasks/nvswitch_assertions.yaml
new file mode 100644
index 0000000..aa131f0
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_assertions.yaml
@@ -0,0 +1,41 @@
+---
+# assertions: FM services active, fabric state, partition count.
+
+- name: Assert FM services are active
+  ansible.builtin.assert:
+    that:
+      - item[1] == 'active'
+    fail_msg: "Service {{ item[0] }} is not active (state: {{ item[1] }})"
+    success_msg: "Service {{ item[0] }} is active"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Assert all GPUs reached Fabric State Completed
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_fabric_states | length == gpu_validation_nvswitch_validation_expected_gpu_count | int
+      - gpu_validation_nvswitch_fabric_completed_count | int == gpu_validation_nvswitch_validation_expected_gpu_count | int
+    fail_msg: >-
+      Discovered {{ gpu_validation_nvswitch_fabric_states | length }} GPU(s), expected {{ gpu_validation_nvswitch_validation_expected_gpu_count }}.
+      Only {{ gpu_validation_nvswitch_fabric_completed_count }} of
+      {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}.
+      States seen: {{ gpu_validation_nvswitch_fabric_states }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}
+  when: gpu_validation_nvswitch_validation_fabric_state_check | bool
+
+- name: Assert expected number of activated partitions
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_activated_partition_count | int == gpu_validation_nvswitch_validation_expected_partition_count | int
+    fail_msg: >-
+      fmpm -l shows {{ gpu_validation_nvswitch_activated_partition_count }} activated partitions,
+      expected {{ gpu_validation_nvswitch_validation_expected_partition_count }}.
+      Full fmpm output:
+      {{ gpu_validation_nvswitch_fmpm_output }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_expected_partition_count }} NVSwitch partitions
+      are activated
+  when:
+    - gpu_validation_nvswitch_validation_expected_partition_count | int > 0
diff --git a/gpu-validation/tasks/nvswitch_isolation.yaml b/gpu-validation/tasks/nvswitch_isolation.yaml
new file mode 100644
index 0000000..0d931e2
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_isolation.yaml
@@ -0,0 +1,43 @@
+---
+# In-VM isolation checks: GPU count, NVLink topology per VM.
+# Runs on the [gpu_vms] group (the passthrough-flavor VMs).
+
+- name: Count GPUs visible inside VM
+  ansible.builtin.command:
+    cmd: nvidia-smi --query-gpu=name --format=csv,noheader
+  register: _vm_gpus
+  changed_when: false
+  failed_when: false
+
+- name: Show GPU list inside VM
+  ansible.builtin.debug:
+    msg: "GPUs in VM {{ inventory_hostname }}: {{ _vm_gpus.stdout_lines }}"
+
+- name: Set GPU count fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_gpu_count: "{{ _vm_gpus.stdout_lines | length }}"
+
+- name: Collect in-VM NVLink topology matrix
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _vm_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show in-VM NVLink topology matrix
+  ansible.builtin.debug:
+    msg: "{{ _vm_topo.stdout | default('nvidia-smi topo not available') }}"
+
+- name: Set in-VM topology facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_topo_output: "{{ _vm_topo.stdout | default('') }}"
+    gpu_validation_nvswitch_vm_nvlink_lines: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('search', '\\bNV[0-9]')
+         | list }}
+    gpu_validation_nvswitch_vm_topo_gpu_count: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('match', '^GPU\\d')
+         | list | length | int }}
diff --git a/gpu-validation/tasks/nvswitch_isolation_assertions.yaml b/gpu-validation/tasks/nvswitch_isolation_assertions.yaml
new file mode 100644
index 0000000..e1e8e2c
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_isolation_assertions.yaml
@@ -0,0 +1,44 @@
+---
+# assertions: GPU count per VM matches flavor, no cross-partition NVLink.
+
+- name: Assert GPU count matches passthrough flavor
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s),
+      expected {{ gpu_validation_nvswitch_validation_gpus_per_vm }} per the passthrough flavor.
+      GPUs seen: {{ _vm_gpus.stdout_lines }}
+    success_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s) as expected
+
+- name: Assert NVLink entries are present (GPUs within partition are connected)
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_nvlink_lines | length > 0
+    fail_msg: >-
+      No NVLink entries found in nvidia-smi topo -m output for
+      VM {{ inventory_hostname }}.  GPUs within the same partition
+      should show NVLink connectivity.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      NVLink connectivity confirmed within partition for
+      VM {{ inventory_hostname }}
+  when:
+    - gpu_validation_nvswitch_validation_gpus_per_vm | int > 1
+
+- name: Assert topo GPU row count matches flavor (no extra GPUs visible from topo matrix)
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_topo_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      nvidia-smi topo -m shows {{ gpu_validation_nvswitch_vm_topo_gpu_count }} GPU row(s) in
+      VM {{ inventory_hostname }} but flavor specifies
+      {{ gpu_validation_nvswitch_validation_gpus_per_vm }}.
+      A mismatch may indicate cross-partition GPU visibility.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      nvidia-smi topo -m GPU row count ({{ gpu_validation_nvswitch_vm_topo_gpu_count }}) matches flavor
+      — no cross-partition GPU leakage detected in VM {{ inventory_hostname }}
diff --git a/gpu-validation/tasks/nvswitch_mode0.yaml b/gpu-validation/tasks/nvswitch_mode0.yaml
new file mode 100644
index 0000000..e3724e0
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode0.yaml
@@ -0,0 +1,89 @@
+---
+# Mode 0 (FABRIC_MODE=0) host-side checks.
+# NVSwitch fabric is open/all-to-all — fmpm is NOT called.
+# FM + NVLSM are required to initialize the fabric; no partition tables are programmed.
+# Isolation is software-only (CUDA + IOMMU). Suitable for trusted tenants only.
+
+- name: Collect FABRIC_MODE from fabricmanager.cfg
+  become: true
+  ansible.builtin.command:
+    cmd: "grep -E '^FABRIC_MODE' {{ gpu_validation_nvswitch_validation_fabric_mode_cfg_file }}"
+  register: _fabric_mode_cfg
+  changed_when: false
+  failed_when: false
+  when: gpu_validation_nvswitch_validation_check_fabric_mode_cfg | bool
+
+- name: Show FABRIC_MODE config value
+  ansible.builtin.debug:
+    msg: "fabricmanager.cfg: {{ _fabric_mode_cfg.stdout | default('not found') }}"
+  when: gpu_validation_nvswitch_validation_check_fabric_mode_cfg | bool
+
+- name: Set FABRIC_MODE config fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fabric_mode_cfg_value: >-
+      {{ (_fabric_mode_cfg.stdout | default('') | regex_search('FABRIC_MODE\s*=\s*(\d+)', '\\1') | default([]) | first | default('-1')) | int }}
+  when: gpu_validation_nvswitch_validation_check_fabric_mode_cfg | bool
+
+- name: Collect FM service status (mode 0)
+  become: true
+  ansible.builtin.systemd:
+    name: "{{ item }}"
+  loop: "{{ gpu_validation_nvswitch_validation_mode0_fm_services }}"
+  register: _fm_service_status
+  failed_when: false
+
+- name: Set FM service status facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fm_service_results: "{{ _fm_service_status.results
+      | map(attribute='item') | zip(
+        _fm_service_status.results
+        | map(attribute='status') | map(attribute='ActiveState'))
+      | list }}"
+
+- name: Show FM service states
+  ansible.builtin.debug:
+    msg: "{{ item[0] }}: {{ item[1] }}"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Collect Fabric State for all GPUs
+  become: true
+  ansible.builtin.shell: |
+    set -o pipefail
+    nvidia-smi -q | grep -A2 'Fabric' | grep 'State' | awk '{print $NF}'
+  args:
+    executable: /bin/bash
+  register: _fabric_states
+  changed_when: false
+  failed_when: false
+
+- name: Show fabric states
+  ansible.builtin.debug:
+    msg: "GPU Fabric States: {{ _fabric_states.stdout_lines }}"
+
+- name: Set fabric state facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fabric_states: "{{ _fabric_states.stdout_lines | map('trim') | list }}"
+    gpu_validation_nvswitch_fabric_completed_count: >-
+      {{ _fabric_states.stdout_lines
+         | map('trim')
+         | select('equalto', gpu_validation_nvswitch_validation_fabric_state_expected)
+         | list | length }}
+
+- name: Collect NVLink topology matrix (open fabric — informational only)
+  become: true
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _nvlink_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show NVLink topology matrix
+  ansible.builtin.debug:
+    msg: |
+      NOTE: Mode 0 NVSwitch fabric is open/all-to-all. All GPUs will show NVLink
+      connectivity to each other. Cross-VM NVLink is NOT blocked at the switch level.
+      {{ _nvlink_topo.stdout | default('nvidia-smi topo not available') }}
+
+- name: Set topology fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_topo_output: "{{ _nvlink_topo.stdout | default('') }}"
diff --git a/gpu-validation/tasks/nvswitch_mode0_assertions.yaml b/gpu-validation/tasks/nvswitch_mode0_assertions.yaml
new file mode 100644
index 0000000..5ec13b6
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode0_assertions.yaml
@@ -0,0 +1,51 @@
+---
+# Mode 0 (FABRIC_MODE=0) host-side assertions.
+# NVSwitch partition check is skipped — fmpm is not used in Mode 0.
+# Isolation is CUDA + IOMMU only; NVSwitch routing is open/all-to-all.
+
+- name: Assert FABRIC_MODE=0 in fabricmanager.cfg
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_fabric_mode_cfg_value | int == gpu_validation_nvswitch_validation_fabric_mode | int
+    fail_msg: >-
+      fabricmanager.cfg has FABRIC_MODE={{ gpu_validation_nvswitch_fabric_mode_cfg_value }}
+      but expected FABRIC_MODE={{ gpu_validation_nvswitch_validation_fabric_mode }}.
+      A leftover mode from a prior deployment will cause incorrect fabric behaviour.
+    success_msg: >-
+      fabricmanager.cfg FABRIC_MODE={{ gpu_validation_nvswitch_fabric_mode_cfg_value }}
+      matches expected mode {{ gpu_validation_nvswitch_validation_fabric_mode }}
+  when:
+    - gpu_validation_nvswitch_validation_check_fabric_mode_cfg | bool
+    - gpu_validation_nvswitch_fabric_mode_cfg_value is defined
+
+- name: Assert FM services are active (mode 0)
+  ansible.builtin.assert:
+    that:
+      - item[1] == 'active'
+    fail_msg: "Service {{ item[0] }} is not active (state: {{ item[1] }})"
+    success_msg: "Service {{ item[0] }} is active"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Assert all GPUs reached Fabric State Completed
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_fabric_states | length == gpu_validation_nvswitch_validation_expected_gpu_count | int
+      - gpu_validation_nvswitch_fabric_completed_count | int == gpu_validation_nvswitch_validation_expected_gpu_count | int
+    fail_msg: >-
+      Discovered {{ gpu_validation_nvswitch_fabric_states | length }} GPU(s), expected {{ gpu_validation_nvswitch_validation_expected_gpu_count }}.
+      Only {{ gpu_validation_nvswitch_fabric_completed_count }} of
+      {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}.
+      States seen: {{ gpu_validation_nvswitch_fabric_states }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}
+  when: gpu_validation_nvswitch_validation_fabric_state_check | bool
+
+- name: Note — NVSwitch partition isolation is not enforced in Mode 0
+  ansible.builtin.debug:
+    msg: >-
+      Mode 0 (FABRIC_MODE=0): NVSwitch routing table is open/all-to-all.
+      fmpm partition check is skipped.
+      NVLink isolation between tenant VMs relies on CUDA device ownership and
+      IOMMU DMA boundaries only. This mode is suitable for trusted tenants only.
diff --git a/gpu-validation/tasks/nvswitch_mode0_isolation.yaml b/gpu-validation/tasks/nvswitch_mode0_isolation.yaml
new file mode 100644
index 0000000..536fb41
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode0_isolation.yaml
@@ -0,0 +1,48 @@
+---
+# Mode 0 (FABRIC_MODE=0) in-VM isolation checks.
+# Runs on the [gpu_vms] group (VMs booted with the PCI-group flavor).
+# Collects GPU count and in-VM NVLink topology.
+# NOTE: NVSwitch routing is open — cross-VM NVLink is NOT blocked at the hardware level.
+
+- name: Count GPUs visible inside VM (mode 0)
+  ansible.builtin.command:
+    cmd: nvidia-smi --query-gpu=name --format=csv,noheader
+  register: _vm_gpus
+  changed_when: false
+  failed_when: false
+
+- name: Show GPU list inside VM
+  ansible.builtin.debug:
+    msg: "GPUs in VM {{ inventory_hostname }}: {{ _vm_gpus.stdout_lines }}"
+
+- name: Set GPU count fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_gpu_count: "{{ _vm_gpus.stdout_lines | length }}"
+
+- name: Collect in-VM NVLink topology matrix
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _vm_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show in-VM NVLink topology matrix
+  ansible.builtin.debug:
+    msg: >-
+      NOTE: Mode 0 open fabric — NVLink visible here is within the assigned PCI group only
+      (the VM cannot access GPUs it does not own via PCI).
+      {{ _vm_topo.stdout | default('nvidia-smi topo not available') }}
+
+- name: Set in-VM topology facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_topo_output: "{{ _vm_topo.stdout | default('') }}"
+    gpu_validation_nvswitch_vm_nvlink_lines: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('search', '\\bNV[0-9]')
+         | list }}
+    gpu_validation_nvswitch_vm_topo_gpu_count: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('match', '^GPU\\d')
+         | list | length | int }}
diff --git a/gpu-validation/tasks/nvswitch_mode0_isolation_assertions.yaml b/gpu-validation/tasks/nvswitch_mode0_isolation_assertions.yaml
new file mode 100644
index 0000000..808525c
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode0_isolation_assertions.yaml
@@ -0,0 +1,36 @@
+---
+# Mode 0 (FABRIC_MODE=0) in-VM isolation assertions.
+# Asserts the VM sees the full PCI group (all GPUs in CUSTOM_GPU_GROUP_N).
+# NVLink cross-VM isolation is NOT asserted — NVSwitch routing is open in Mode 0.
+
+- name: Assert GPU count matches PCI group size
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s),
+      expected {{ gpu_validation_nvswitch_validation_gpus_per_vm }} (full PCI group size).
+      GPUs seen: {{ _vm_gpus.stdout_lines }}
+    success_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s) — full PCI group assigned
+
+- name: Assert NVLink entries are present within the PCI group
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_nvlink_lines | length > 0
+    fail_msg: >-
+      No NVLink entries found in nvidia-smi topo -m for VM {{ inventory_hostname }}.
+      GPUs within the same PCI group should show NVLink connectivity.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      NVLink connectivity confirmed within PCI group for VM {{ inventory_hostname }}
+  when:
+    - gpu_validation_nvswitch_validation_gpus_per_vm | int > 1
+
+- name: Note — cross-VM NVLink isolation is software-only in Mode 0
+  ansible.builtin.debug:
+    msg: >-
+      Mode 0 (FABRIC_MODE=0): NVSwitch routing table is open/all-to-all.
+      Cross-VM NVLink hardware isolation is NOT enforced. This mode is
+      suitable for trusted tenants only. For hardware isolation use Mode 1 or Mode 2.
diff --git a/gpu-validation/tasks/nvswitch_mode1.yaml b/gpu-validation/tasks/nvswitch_mode1.yaml
new file mode 100644
index 0000000..0925c7c
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode1.yaml
@@ -0,0 +1,84 @@
+---
+# host-side checks: FM services, fabric state, partition activation.
+
+- name: Collect FM service status
+  become: true
+  ansible.builtin.systemd:
+    name: "{{ item }}"
+  loop: "{{ gpu_validation_nvswitch_validation_mode1_fm_services }}"
+  register: _fm_service_status
+  failed_when: false
+
+- name: Set FM service status facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fm_service_results: "{{ _fm_service_status.results
+      | map(attribute='item') | zip(
+        _fm_service_status.results
+        | map(attribute='status') | map(attribute='ActiveState'))
+      | list }}"
+
+- name: Show FM service states
+  ansible.builtin.debug:
+    msg: "{{ item[0] }}: {{ item[1] }}"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Collect Fabric State for all GPUs
+  become: true
+  ansible.builtin.shell: |
+    set -o pipefail
+    nvidia-smi -q | grep -A2 'Fabric' | grep 'State' | awk '{print $NF}'
+  args:
+    executable: /bin/bash
+  register: _fabric_states
+  changed_when: false
+  failed_when: false
+
+- name: Show fabric states
+  ansible.builtin.debug:
+    msg: "GPU Fabric States: {{ _fabric_states.stdout_lines }}"
+
+- name: Set fabric state facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fabric_states: "{{ _fabric_states.stdout_lines | map('trim') | list }}"
+    gpu_validation_nvswitch_fabric_completed_count: >-
+      {{ _fabric_states.stdout_lines
+         | map('trim')
+         | select('equalto', gpu_validation_nvswitch_validation_fabric_state_expected)
+         | list | length }}
+
+- name: Collect fmpm partition list
+  become: true
+  ansible.builtin.command:
+    cmd: fmpm -l
+  register: _fmpm_list
+  changed_when: false
+  failed_when: false
+
+- name: Show fmpm partition list
+  ansible.builtin.debug:
+    msg: "{{ _fmpm_list.stdout | default('fmpm not available or no output') }}"
+
+- name: Set partition facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fmpm_output: "{{ _fmpm_list.stdout | default('') }}"
+    gpu_validation_nvswitch_activated_partition_count: >-
+      {{ (_fmpm_list.stdout | default(''))
+         .splitlines()
+         | select('search', '\\bactivated\\b')
+         | list | length }}
+
+- name: Collect NVLink topology matrix
+  become: true
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _nvlink_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show NVLink topology matrix
+  ansible.builtin.debug:
+    msg: "{{ _nvlink_topo.stdout | default('nvidia-smi topo not available') }}"
+
+- name: Set topology fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_topo_output: "{{ _nvlink_topo.stdout | default('') }}"
diff --git a/gpu-validation/tasks/nvswitch_mode1_assertions.yaml b/gpu-validation/tasks/nvswitch_mode1_assertions.yaml
new file mode 100644
index 0000000..aa131f0
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode1_assertions.yaml
@@ -0,0 +1,41 @@
+---
+# assertions: FM services active, fabric state, partition count.
+
+- name: Assert FM services are active
+  ansible.builtin.assert:
+    that:
+      - item[1] == 'active'
+    fail_msg: "Service {{ item[0] }} is not active (state: {{ item[1] }})"
+    success_msg: "Service {{ item[0] }} is active"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Assert all GPUs reached Fabric State Completed
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_fabric_states | length == gpu_validation_nvswitch_validation_expected_gpu_count | int
+      - gpu_validation_nvswitch_fabric_completed_count | int == gpu_validation_nvswitch_validation_expected_gpu_count | int
+    fail_msg: >-
+      Discovered {{ gpu_validation_nvswitch_fabric_states | length }} GPU(s), expected {{ gpu_validation_nvswitch_validation_expected_gpu_count }}.
+      Only {{ gpu_validation_nvswitch_fabric_completed_count }} of
+      {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}.
+      States seen: {{ gpu_validation_nvswitch_fabric_states }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}
+  when: gpu_validation_nvswitch_validation_fabric_state_check | bool
+
+- name: Assert expected number of activated partitions
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_activated_partition_count | int == gpu_validation_nvswitch_validation_expected_partition_count | int
+    fail_msg: >-
+      fmpm -l shows {{ gpu_validation_nvswitch_activated_partition_count }} activated partitions,
+      expected {{ gpu_validation_nvswitch_validation_expected_partition_count }}.
+      Full fmpm output:
+      {{ gpu_validation_nvswitch_fmpm_output }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_expected_partition_count }} NVSwitch partitions
+      are activated
+  when:
+    - gpu_validation_nvswitch_validation_expected_partition_count | int > 0
diff --git a/gpu-validation/tasks/nvswitch_mode1_isolation.yaml b/gpu-validation/tasks/nvswitch_mode1_isolation.yaml
new file mode 100644
index 0000000..0d931e2
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode1_isolation.yaml
@@ -0,0 +1,43 @@
+---
+# In-VM isolation checks: GPU count, NVLink topology per VM.
+# Runs on the [gpu_vms] group (the passthrough-flavor VMs).
+
+- name: Count GPUs visible inside VM
+  ansible.builtin.command:
+    cmd: nvidia-smi --query-gpu=name --format=csv,noheader
+  register: _vm_gpus
+  changed_when: false
+  failed_when: false
+
+- name: Show GPU list inside VM
+  ansible.builtin.debug:
+    msg: "GPUs in VM {{ inventory_hostname }}: {{ _vm_gpus.stdout_lines }}"
+
+- name: Set GPU count fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_gpu_count: "{{ _vm_gpus.stdout_lines | length }}"
+
+- name: Collect in-VM NVLink topology matrix
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _vm_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show in-VM NVLink topology matrix
+  ansible.builtin.debug:
+    msg: "{{ _vm_topo.stdout | default('nvidia-smi topo not available') }}"
+
+- name: Set in-VM topology facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_topo_output: "{{ _vm_topo.stdout | default('') }}"
+    gpu_validation_nvswitch_vm_nvlink_lines: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('search', '\\bNV[0-9]')
+         | list }}
+    gpu_validation_nvswitch_vm_topo_gpu_count: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('match', '^GPU\\d')
+         | list | length | int }}
diff --git a/gpu-validation/tasks/nvswitch_mode1_isolation_assertions.yaml b/gpu-validation/tasks/nvswitch_mode1_isolation_assertions.yaml
new file mode 100644
index 0000000..e1e8e2c
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode1_isolation_assertions.yaml
@@ -0,0 +1,44 @@
+---
+# assertions: GPU count per VM matches flavor, no cross-partition NVLink.
+
+- name: Assert GPU count matches passthrough flavor
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s),
+      expected {{ gpu_validation_nvswitch_validation_gpus_per_vm }} per the passthrough flavor.
+      GPUs seen: {{ _vm_gpus.stdout_lines }}
+    success_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s) as expected
+
+- name: Assert NVLink entries are present (GPUs within partition are connected)
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_nvlink_lines | length > 0
+    fail_msg: >-
+      No NVLink entries found in nvidia-smi topo -m output for
+      VM {{ inventory_hostname }}.  GPUs within the same partition
+      should show NVLink connectivity.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      NVLink connectivity confirmed within partition for
+      VM {{ inventory_hostname }}
+  when:
+    - gpu_validation_nvswitch_validation_gpus_per_vm | int > 1
+
+- name: Assert topo GPU row count matches flavor (no extra GPUs visible from topo matrix)
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_topo_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      nvidia-smi topo -m shows {{ gpu_validation_nvswitch_vm_topo_gpu_count }} GPU row(s) in
+      VM {{ inventory_hostname }} but flavor specifies
+      {{ gpu_validation_nvswitch_validation_gpus_per_vm }}.
+      A mismatch may indicate cross-partition GPU visibility.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      nvidia-smi topo -m GPU row count ({{ gpu_validation_nvswitch_vm_topo_gpu_count }}) matches flavor
+      — no cross-partition GPU leakage detected in VM {{ inventory_hostname }}
diff --git a/gpu-validation/tasks/nvswitch_mode2.yaml b/gpu-validation/tasks/nvswitch_mode2.yaml
new file mode 100644
index 0000000..c8c8f91
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode2.yaml
@@ -0,0 +1,82 @@
+---
+# Mode 2 (FABRIC_MODE=2) host-side checks.
+# nvidia-vgpu-mgr programs NVSwitch partition routing automatically per mdev instance.
+# mdevctl list verifies one mdev instance per GPU PF BDF.
+
+- name: Collect FM service status (mode 2)
+  become: true
+  ansible.builtin.systemd:
+    name: "{{ item }}"
+  loop: "{{ gpu_validation_nvswitch_validation_mode2_fm_services }}"
+  register: _fm_service_status
+  failed_when: false
+
+- name: Set FM service status facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fm_service_results: "{{ _fm_service_status.results
+      | map(attribute='item') | zip(
+         _fm_service_status.results
+         | map(attribute='status') | map(attribute='ActiveState'))
+      | list }}"
+
+- name: Show FM service states
+  ansible.builtin.debug:
+    msg: "{{ item[0] }}: {{ item[1] }}"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Collect Fabric State for all GPUs
+  become: true
+  ansible.builtin.shell: |
+    set -o pipefail
+    nvidia-smi -q | grep -A2 'Fabric' | grep 'State' | awk '{print $NF}'
+  args:
+    executable: /bin/bash
+  register: _fabric_states
+  changed_when: false
+  failed_when: false
+
+- name: Show fabric states
+  ansible.builtin.debug:
+    msg: "GPU Fabric States: {{ _fabric_states.stdout_lines }}"
+
+- name: Set fabric state facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_fabric_states: "{{ _fabric_states.stdout_lines | map('trim') | list }}"
+    gpu_validation_nvswitch_fabric_completed_count: >-
+      {{ _fabric_states.stdout_lines
+         | map('trim')
+         | select('equalto', gpu_validation_nvswitch_validation_fabric_state_expected)
+         | list | length }}
+
+- name: Collect mdev instance list
+  become: true
+  ansible.builtin.command:
+    cmd: mdevctl list
+  register: _mdevctl_list
+  changed_when: false
+  failed_when: false
+
+- name: Show mdev instance list
+  ansible.builtin.debug:
+    msg: "{{ _mdevctl_list.stdout | default('mdevctl list returned no output') }}"
+
+- name: Set mdev instance count fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_mdev_instance_count: "{{ _mdevctl_list.stdout_lines | default([]) | length }}"
+    gpu_validation_nvswitch_mdevctl_output: "{{ _mdevctl_list.stdout | default('') }}"
+
+- name: Collect NVLink topology matrix
+  become: true
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _nvlink_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show NVLink topology matrix
+  ansible.builtin.debug:
+    msg: "{{ _nvlink_topo.stdout | default('nvidia-smi topo not available') }}"
+
+- name: Set topology fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_topo_output: "{{ _nvlink_topo.stdout | default('') }}"
diff --git a/gpu-validation/tasks/nvswitch_mode2_assertions.yaml b/gpu-validation/tasks/nvswitch_mode2_assertions.yaml
new file mode 100644
index 0000000..6fd01ca
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode2_assertions.yaml
@@ -0,0 +1,40 @@
+---
+# Mode 2 (FABRIC_MODE=2) host-side assertions.
+# Verifies FM + nvidia-vgpu-mgr are active, all GPUs completed, and mdev instances exist.
+
+- name: Assert FM services are active (mode 2)
+  ansible.builtin.assert:
+    that:
+      - item[1] == 'active'
+    fail_msg: "Service {{ item[0] }} is not active (state: {{ item[1] }})"
+    success_msg: "Service {{ item[0] }} is active"
+  loop: "{{ gpu_validation_nvswitch_fm_service_results }}"
+
+- name: Assert all GPUs reached Fabric State Completed
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_fabric_states | length == gpu_validation_nvswitch_validation_expected_gpu_count | int
+      - gpu_validation_nvswitch_fabric_completed_count | int == gpu_validation_nvswitch_validation_expected_gpu_count | int
+    fail_msg: >-
+      Discovered {{ gpu_validation_nvswitch_fabric_states | length }} GPU(s), expected {{ gpu_validation_nvswitch_validation_expected_gpu_count }}.
+      Only {{ gpu_validation_nvswitch_fabric_completed_count }} of
+      {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}.
+      States seen: {{ gpu_validation_nvswitch_fabric_states }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_expected_gpu_count }} GPUs reached
+      Fabric State: {{ gpu_validation_nvswitch_validation_fabric_state_expected }}
+  when: gpu_validation_nvswitch_validation_fabric_state_check | bool
+
+- name: Assert expected number of mdev instances exist
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_mdev_instance_count | int == gpu_validation_nvswitch_validation_mode2_expected_mdev_count | int
+    fail_msg: >-
+      mdevctl list shows {{ gpu_validation_nvswitch_mdev_instance_count }} mdev instance(s),
+      expected {{ gpu_validation_nvswitch_validation_mode2_expected_mdev_count }}.
+      Full mdevctl output:
+      {{ gpu_validation_nvswitch_mdevctl_output }}
+    success_msg: >-
+      All {{ gpu_validation_nvswitch_validation_mode2_expected_mdev_count }} mdev instance(s) found
+  when: gpu_validation_nvswitch_validation_mode2_expected_mdev_count | int > 0
diff --git a/gpu-validation/tasks/nvswitch_mode2_isolation.yaml b/gpu-validation/tasks/nvswitch_mode2_isolation.yaml
new file mode 100644
index 0000000..2dbebe9
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode2_isolation.yaml
@@ -0,0 +1,45 @@
+---
+# Mode 2 (FABRIC_MODE=2) in-VM isolation checks.
+# Runs on the [gpu_vms] group (VMs booted with the mdev flavor).
+# From inside the VM the mdev device presents as the physical GPU model;
+# nvidia-smi output is equivalent to Mode 1 passthrough from the guest's perspective.
+
+- name: Count GPUs visible inside VM (mode 2)
+  ansible.builtin.command:
+    cmd: nvidia-smi --query-gpu=name --format=csv,noheader
+  register: _vm_gpus
+  changed_when: false
+  failed_when: false
+
+- name: Show GPU list inside VM
+  ansible.builtin.debug:
+    msg: "GPUs in VM {{ inventory_hostname }}: {{ _vm_gpus.stdout_lines }}"
+
+- name: Set GPU count fact
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_gpu_count: "{{ _vm_gpus.stdout_lines | length }}"
+
+- name: Collect in-VM NVLink topology matrix
+  ansible.builtin.command:
+    cmd: nvidia-smi topo -m
+  register: _vm_topo
+  changed_when: false
+  failed_when: false
+
+- name: Show in-VM NVLink topology matrix
+  ansible.builtin.debug:
+    msg: "{{ _vm_topo.stdout | default('nvidia-smi topo not available') }}"
+
+- name: Set in-VM topology facts
+  ansible.builtin.set_fact:
+    gpu_validation_nvswitch_vm_topo_output: "{{ _vm_topo.stdout | default('') }}"
+    gpu_validation_nvswitch_vm_nvlink_lines: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('search', '\\bNV[0-9]')
+         | list }}
+    gpu_validation_nvswitch_vm_topo_gpu_count: >-
+      {{ (_vm_topo.stdout | default(''))
+         .splitlines()
+         | select('match', '^GPU\\d')
+         | list | length | int }}
diff --git a/gpu-validation/tasks/nvswitch_mode2_isolation_assertions.yaml b/gpu-validation/tasks/nvswitch_mode2_isolation_assertions.yaml
new file mode 100644
index 0000000..d3c4857
--- /dev/null
+++ b/gpu-validation/tasks/nvswitch_mode2_isolation_assertions.yaml
@@ -0,0 +1,43 @@
+---
+# Mode 2 (FABRIC_MODE=2) in-VM isolation assertions.
+# nvidia-vgpu-mgr programs NVSwitch partition routing per mdev — hardware-enforced isolation.
+
+- name: Assert GPU count matches mdev flavor slot size
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s),
+      expected {{ gpu_validation_nvswitch_validation_gpus_per_vm }} per the mdev flavor.
+      GPUs seen: {{ _vm_gpus.stdout_lines }}
+    success_msg: >-
+      VM {{ inventory_hostname }} has {{ gpu_validation_nvswitch_vm_gpu_count }} GPU(s) as expected
+
+- name: Assert NVLink entries are present within the mdev partition
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_nvlink_lines | length > 0
+    fail_msg: >-
+      No NVLink entries found in nvidia-smi topo -m for VM {{ inventory_hostname }}.
+      GPUs within the same mdev partition should show NVLink connectivity.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      NVLink connectivity confirmed within mdev partition for VM {{ inventory_hostname }}
+  when:
+    - gpu_validation_nvswitch_validation_gpus_per_vm | int > 1
+
+- name: Assert topo GPU row count matches mdev slot (no extra GPUs visible from topo matrix)
+  ansible.builtin.assert:
+    that:
+      - gpu_validation_nvswitch_vm_topo_gpu_count | int == gpu_validation_nvswitch_validation_gpus_per_vm | int
+    fail_msg: >-
+      nvidia-smi topo -m shows {{ gpu_validation_nvswitch_vm_topo_gpu_count }} GPU row(s) in
+      VM {{ inventory_hostname }} but mdev flavor specifies
+      {{ gpu_validation_nvswitch_validation_gpus_per_vm }}.
+      A mismatch may indicate cross-partition GPU visibility.
+      Topology output:
+      {{ gpu_validation_nvswitch_vm_topo_output }}
+    success_msg: >-
+      nvidia-smi topo -m GPU row count ({{ gpu_validation_nvswitch_vm_topo_gpu_count }}) matches mdev flavor
+      — no cross-partition GPU leakage detected in VM {{ inventory_hostname }}
diff --git a/nvswitch_validation.yaml b/nvswitch_validation.yaml
new file mode 100644
index 0000000..40794f0
--- /dev/null
+++ b/nvswitch_validation.yaml
@@ -0,0 +1,112 @@
+---
+# NVSwitch validation playbook — dispatches to mode-specific task files.
+#
+# REQUIRED: gpu_validation_nvswitch_validation_fabric_mode must be supplied by the caller.
+# It cannot come from role defaults because tasks_from is rendered in the play
+# context before role defaults are loaded.  Set it via:
+#   - extra-vars:  -e gpu_validation_nvswitch_validation_fabric_mode=0   ← preferred (single source)
+#   - group_vars/all                                       ← also single source
+#   - AnsibleTest CR ansibleVarFiles
+#
+# WARNING: do NOT set it in group_vars/edpm_hosts or group_vars/gpu_vms separately.
+# The two plays resolve the variable independently from their own host context; a
+# per-group assignment silently dispatches each play to a different mode task file.
+#
+# Valid values:
+#   0 — PCI Grouping (FABRIC_MODE=0, open NVSwitch fabric, trusted tenants only)
+#   1 — Shared NVSwitch (FABRIC_MODE=1, fmpm partitions, hardware isolation)
+#   2 — vGPU Multitenancy (FABRIC_MODE=2, mdev, hardware isolation)
+#
+# Host-side checks (runs on edpm_hosts):
+#   Mode 0: FM + NVLSM active, all GPUs Fabric State Completed, open-fabric topo display
+#   Mode 1: FM + NVLSM + fm-partition-restore active, fmpm -l partitions activated
+#   Mode 2: FM + NVLSM + nvidia-vgpu-mgr active, mdevctl list count matches expected
+#
+# In-VM isolation checks (runs on gpu_vms):
+#   Mode 0: GPU count matches PCI group size, NVLink within group (no switch-level isolation)
+#   Mode 1: GPU count matches partition size, NVLink within partition (hardware isolation)
+#   Mode 2: GPU count matches mdev slot size, NVLink within mdev partition (hardware isolation)
+#
+# Inventory groups required:
+#   [edpm_hosts]  — the EDPM compute node(s)
+#   [gpu_vms]     — test VMs booted with the GPU flavor (two minimum)
+#
+# See gpu-validation/defaults/main.yaml for all other gpu_validation_nvswitch_validation_* defaults.
+
+- name: Pre-flight — validate inventory groups
+  hosts: localhost
+  gather_subset:
+    - min
+  tasks:
+    - name: Assert edpm_hosts is non-empty
+      ansible.builtin.assert:
+        that: groups['edpm_hosts'] | default([]) | length >= 1
+        fail_msg: >-
+          Inventory group [edpm_hosts] is empty or undefined.
+          Add at least one EDPM compute node to this group.
+
+    - name: Assert gpu_vms has at least two hosts
+      ansible.builtin.assert:
+        that: groups['gpu_vms'] | default([]) | length >= 2
+        fail_msg: >-
+          Inventory group [gpu_vms] has {{ groups['gpu_vms'] | default([]) | length }} host(s);
+          at least two are required to validate cross-VM NVSwitch isolation.
+
+- name: Host-side checks
+  hosts: edpm_hosts
+  gather_subset:
+    - min
+  pre_tasks:
+    - name: Assert gpu_validation_nvswitch_validation_fabric_mode is provided by caller
+      ansible.builtin.assert:
+        that:
+          - gpu_validation_nvswitch_validation_fabric_mode is defined
+          - (gpu_validation_nvswitch_validation_fabric_mode | string) in ['0', '1', '2']
+        fail_msg: >-
+          gpu_validation_nvswitch_validation_fabric_mode must be set to 0, 1, or 2 via
+          extra-vars (-e gpu_validation_nvswitch_validation_fabric_mode=N), inventory/group_vars,
+          or an AnsibleTest CR ansibleVarFiles entry.
+          It cannot come from role defaults because tasks_from is resolved in the
+          play context before role defaults are loaded.
+  tasks:
+    - name: "Collect host-side data (mode {{ gpu_validation_nvswitch_validation_fabric_mode }})"
+      ansible.builtin.include_role:
+        name: gpu-validation
+        tasks_from: "nvswitch_mode{{ gpu_validation_nvswitch_validation_fabric_mode }}.yaml"
+
+    - name: "Assert host-side state (mode {{ gpu_validation_nvswitch_validation_fabric_mode }})"
+      ansible.builtin.include_role:
+        name: gpu-validation
+        tasks_from: "nvswitch_mode{{ gpu_validation_nvswitch_validation_fabric_mode }}_assertions.yaml"
+
+- name: In-VM isolation checks
+  hosts: gpu_vms
+  gather_subset:
+    - min
+  pre_tasks:
+    - name: Assert fabric mode is valid and consistent with edpm_hosts
+      ansible.builtin.assert:
+        that:
+          - gpu_validation_nvswitch_validation_fabric_mode is defined
+          - (gpu_validation_nvswitch_validation_fabric_mode | string) in ['0', '1', '2']
+          - groups['edpm_hosts'] | map('extract', hostvars, 'gpu_validation_nvswitch_validation_fabric_mode')
+            | list | unique | length == 1
+          - (groups['edpm_hosts'] | map('extract', hostvars, 'gpu_validation_nvswitch_validation_fabric_mode')
+            | list | first | string) == (gpu_validation_nvswitch_validation_fabric_mode | string)
+        fail_msg: >-
+          gpu_validation_nvswitch_validation_fabric_mode is not consistent across inventory groups.
+          gpu_vms resolved {{ gpu_validation_nvswitch_validation_fabric_mode }};
+          edpm_hosts resolved
+          {{ groups['edpm_hosts'] | map('extract', hostvars, 'gpu_validation_nvswitch_validation_fabric_mode') | list }}.
+          Set the variable via extra-vars or group_vars/all (not group-specific vars)
+          so both plays dispatch to the same mode task file.
+  tasks:
+    - name: "Collect in-VM data (mode {{ gpu_validation_nvswitch_validation_fabric_mode }})"
+      ansible.builtin.include_role:
+        name: gpu-validation
+        tasks_from: "nvswitch_mode{{ gpu_validation_nvswitch_validation_fabric_mode }}_isolation.yaml"
+
+    - name: "Assert in-VM isolation (mode {{ gpu_validation_nvswitch_validation_fabric_mode }})"
+      ansible.builtin.include_role:
+        name: gpu-validation
+        tasks_from: "nvswitch_mode{{ gpu_validation_nvswitch_validation_fabric_mode }}_isolation_assertions.yaml"