diff --git a/.buildkite/test-amd.yaml b/.buildkite/test-amd.yaml
index fe2cfd18c8..e0c0812866 100644
--- a/.buildkite/test-amd.yaml
+++ b/.buildkite/test-amd.yaml
@@ -436,7 +436,7 @@ steps:
   - tests/v1/e2e/general/
   - vllm/platforms/rocm.py
   commands:
-  - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
+  - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py --ignore v1/e2e/general/test_kv_sharing_fast_prefill.py -k "not test_mamba_prefix_cache_mrv2"
 
 - label: ":amd: (MI250) E2E Scheduling" # TBD
   timeout_in_minutes: 55
@@ -1737,29 +1737,6 @@ steps:
   - pytest -v -s tests/ir
   - pytest -v -s tests/kernels/ir
 
-- label: ":amd: (MI300) MLA Kernels" # TBD
-  timeout_in_minutes: 25
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  dind: false
-  agent_pool: mi300_1
-  working_dir: "/vllm-workspace/"
-  source_file_dependencies:
-  - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
-  - vllm/v1/attention/selector.py
-  - vllm/platforms/rocm.py
-  - vllm/_aiter_ops.py
-  - tests/kernels/attention/test_rocm_aiter_mla_decode.py
-  - tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
-  - tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
-  - tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
-  optional: true
-  commands:
-  - amd-smi
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
-
 - label: ":amd: (MI300) Attention Kernels Shard %N" # TBD
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -1778,7 +1755,7 @@ steps:
   - vllm/platforms/rocm.py
   - vllm/utils/flashinfer.py
   commands:
-  - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  - pytest -v -s kernels/attention --ignore=kernels/attention/test_triton_unified_attention_diffkv.py --ignore=kernels/attention/test_rocm_aiter_mla_decode.py --ignore=kernels/attention/test_rocm_aiter_mla_decode_metadata.py --ignore=kernels/attention/test_rocm_aiter_mla_fp8_support.py --ignore=kernels/attention/test_rocm_aiter_mla_op_registration.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
 - label: ":amd: (MI300) MiniMax Reduce RMS Kernels" # TBD
   timeout_in_minutes: 45
@@ -2566,7 +2543,7 @@ steps:
   - vllm/platforms/rocm.py
   - vllm/entrypoints/generate/beam_search
   commands:
-  - pytest -v -s samplers
+  - pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
 
 #------------------------------------------------------------  mi300 · misc  ------------------------------------------------------------#
 
@@ -3364,6 +3341,48 @@ steps:
   commands:
   - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k"
 
+#----------------------------------------------------------  mi355 · compile  ----------------------------------------------------------#
+
+- label: ":amd: (MI355) PyTorch Compilation H100 Cases"
+  timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
+  agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/__init__.py
+  - vllm/_aiter_ops.py
+  - vllm/_custom_ops.py
+  - vllm/compilation/
+  - vllm/config/
+  - vllm/distributed/
+  - "!vllm/distributed/kv_transfer/"
+  - vllm/engine/
+  - vllm/env_override.py
+  - vllm/envs.py
+  - vllm/forward_context.py
+  - vllm/inputs/
+  - vllm/ir/
+  - vllm/kernels/
+  - vllm/logger.py
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/plugins/
+  - vllm/sampling_params.py
+  - vllm/sequence.py
+  - vllm/transformers_utils/
+  - vllm/triton_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/compile/h100/
+  - tests/utils.py
+  - csrc/
+  commands:
+  - pytest -v -s compile/h100/test_startup.py
+
 #--------------------------------------------------------  mi355 · distributed  --------------------------------------------------------#
 
 - label: ":amd: (MI355) Distributed DP Basic" # TBD
@@ -3790,6 +3809,23 @@ steps:
 
 #----------------------------------------------------------  mi355 · kernels  ----------------------------------------------------------#
 
+- label: ":amd: (MI355) Attention DiffKV Kernels"
+  timeout_in_minutes: 35
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
+  agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/attention/ops/triton_unified_attention_diffkv.py
+  - vllm/v1/attention/backends/triton_attn_diffkv.py
+  - tests/kernels/attention/test_triton_unified_attention.py
+  - tests/kernels/attention/test_triton_unified_attention_diffkv.py
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s kernels/attention/test_triton_unified_attention_diffkv.py
+
 - label: ":amd: (MI355) Core Operation Kernels Shard %N" # TBD
   timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -3859,19 +3895,19 @@ steps:
   - vllm/_aiter_ops.py
   - examples/basic/offline_inference/chat.py
   - tests/kernels/attention/test_attention_selector.py
-  - tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
   optional: true
   commands:
   - amd-smi
   - python3 examples/basic/offline_inference/chat.py --attention-backend TRITON_ATTN
   - pytest -v -s tests/kernels/attention/test_attention_selector.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
 
-- label: ":amd: (MI355) MLA Kernels" # TBD
-  timeout_in_minutes: 25
+- label: ":amd: (MI355) MLA Kernels"
+  timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   dind: false
   agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
   working_dir: "/vllm-workspace/"
   source_file_dependencies:
   - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
@@ -3882,7 +3918,6 @@ steps:
   - tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
   - tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
   - tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
-  optional: true
   commands:
   - amd-smi
   - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
@@ -3908,7 +3943,7 @@ steps:
   - vllm/utils/flashinfer.py
   optional: true
   commands:
-  - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  - pytest -v -s kernels/attention --ignore=kernels/attention/test_triton_unified_attention_diffkv.py --ignore=kernels/attention/test_rocm_aiter_mla_decode.py --ignore=kernels/attention/test_rocm_aiter_mla_decode_metadata.py --ignore=kernels/attention/test_rocm_aiter_mla_fp8_support.py --ignore=kernels/attention/test_rocm_aiter_mla_op_registration.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
 - label: ":amd: (MI355) MoE Kernels Shard %N" # TBD
   timeout_in_minutes: 55
@@ -4050,6 +4085,21 @@ steps:
 
 #------------------------------------------------------  mi355 · models / basic  -------------------------------------------------------#
 
+- label: ":amd: (MI355) GLM5Next"
+  timeout_in_minutes: 35
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
+  agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/models/glm5next/
+  - tests/models/glm5next/
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s models/glm5next
+
 - label: ":amd: (MI355) Basic Models (Extra Initialization) Shard %N" # TBD
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -4087,7 +4137,27 @@ steps:
   - export GIT_CONFIG_VALUE_0=HTTP/1.1
   - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
   - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-  - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+  - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
+- label: ":amd: (MI355) Granite Language Model Compatibility"
+  timeout_in_minutes: 80
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
+  agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - "!vllm/distributed/kv_transfer/"
+  - tests/models/language/generation
+  commands:
+  - export GIT_CONFIG_COUNT=1
+  - export GIT_CONFIG_KEY_0=http.version
+  - export GIT_CONFIG_VALUE_0=HTTP/1.1
+  - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
+  - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
+  - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
 
 - label: ":amd: (MI355) Language Models (Extended Generation)" # TBD
   timeout_in_minutes: 95
@@ -4308,8 +4378,46 @@ steps:
 #   commands:
 #   - pytest -s -v tests/quantization/test_gfx950_moe.py
 
+#---------------------------------------------------------  mi355 · samplers  ----------------------------------------------------------#
+
+- label: ":amd: (MI355) Samplers Multimodal Beam Search"
+  timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
+  agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/sampling_metadata.py
+  - vllm/v1/sample/
+  - vllm/entrypoints/generate/beam_search/
+  - tests/samplers
+  - tests/conftest.py
+  - vllm/_aiter_ops.py
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
+
 #------------------------------------------------------------  mi355 · v1  -------------------------------------------------------------#
 
+- label: ":amd: (MI355) E2E Core Large Memory"
+  timeout_in_minutes: 65
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
+  agent_pool: mi355_1
+  num_gpus: 1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/
+  - tests/v1/e2e/general/
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s v1/e2e/general/test_kv_sharing_fast_prefill.py
+  - pytest -v -s v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2 v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2_async
+
 - label: ":amd: (MI355) V1 Attention Shard %N" # TBD
   timeout_in_minutes: 125
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
diff --git a/.buildkite/test_areas/engine.yaml b/.buildkite/test_areas/engine.yaml
index 8e47a84e18..e38cd92573 100644
--- a/.buildkite/test_areas/engine.yaml
+++ b/.buildkite/test_areas/engine.yaml
@@ -94,7 +94,7 @@ steps:
       label: ":amd: (MI250) E2E Core"
       dind: false
       commands:
-      - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
+      - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py --ignore v1/e2e/general/test_kv_sharing_fast_prefill.py -k "not test_mamba_prefix_cache_mrv2"
       device: mi250_1
       timeout_in_minutes: 50
       depends_on:
@@ -115,6 +115,23 @@ steps:
   commands:
     - pytest -v -s v1/e2e/general/test_kv_sharing_fast_prefill.py
     - pytest -v -s v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2 v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2_async
+  mirror:
+    amd:
+      label: ":amd: (MI355) E2E Core Large Memory"
+      dind: false
+      device: mi355_1
+      num_devices: 1
+      timeout_in_minutes: 65
+      working_dir: /vllm-workspace/tests
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/
+      - tests/v1/e2e/general/
+      - vllm/platforms/rocm.py
+      commands:
+      - pytest -v -s v1/e2e/general/test_kv_sharing_fast_prefill.py
+      - pytest -v -s v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2 v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2_async
 
 - label: ":nvidia: (L4) V1 E2E"
   key: v1-e2e-2-gpus
diff --git a/.buildkite/test_areas/kernels.yaml b/.buildkite/test_areas/kernels.yaml
index aab6e52b8e..2134d6dd04 100644
--- a/.buildkite/test_areas/kernels.yaml
+++ b/.buildkite/test_areas/kernels.yaml
@@ -196,6 +196,8 @@ steps:
       - vllm/_aiter_ops.py
       - vllm/envs.py
       - vllm/platforms/rocm.py
+      commands:
+      - pytest -v -s kernels/attention --ignore=kernels/attention/test_triton_unified_attention_diffkv.py --ignore=kernels/attention/test_rocm_aiter_mla_decode.py --ignore=kernels/attention/test_rocm_aiter_mla_decode_metadata.py --ignore=kernels/attention/test_rocm_aiter_mla_fp8_support.py --ignore=kernels/attention/test_rocm_aiter_mla_op_registration.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
 - label: ":nvidia: (H100) Attention DiffKV Kernels"
   key: kernels-attention-diffkv-test-h100
@@ -209,6 +211,24 @@ steps:
   - tests/kernels/attention/test_triton_unified_attention_diffkv.py
   commands:
     - pytest -v -s kernels/attention/test_triton_unified_attention_diffkv.py
+  mirror:
+    amd:
+      label: ":amd: (MI355) Attention DiffKV Kernels"
+      dind: false
+      device: mi355_1
+      num_devices: 1
+      timeout_in_minutes: 35
+      working_dir: /vllm-workspace/tests
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/attention/ops/triton_unified_attention_diffkv.py
+      - vllm/v1/attention/backends/triton_attn_diffkv.py
+      - tests/kernels/attention/test_triton_unified_attention.py
+      - tests/kernels/attention/test_triton_unified_attention_diffkv.py
+      - vllm/platforms/rocm.py
+      commands:
+      - pytest -v -s kernels/attention/test_triton_unified_attention_diffkv.py
 
 - label: ":nvidia: (H100) MLA Kernel Test"
   key: kernels-flashmla-test-h100
@@ -231,6 +251,31 @@ steps:
     - pytest -v -s kernels/attention/test_flashmla_sparse.py
     - pytest -v -s kernels/attention/test_flashinfer_mla_decode.py
     - pytest -v -s kernels/attention/test_mla_cross_layer_kernel_equivalence.py
+  mirror:
+    amd:
+      label: ":amd: (MI355) MLA Kernels"
+      dind: false
+      device: mi355_1
+      num_devices: 1
+      timeout_in_minutes: 40
+      working_dir: /vllm-workspace/
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
+      - vllm/v1/attention/selector.py
+      - vllm/platforms/rocm.py
+      - vllm/_aiter_ops.py
+      - tests/kernels/attention/test_rocm_aiter_mla_decode.py
+      - tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
+      - tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
+      - tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
+      commands:
+      - amd-smi
+      - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
+      - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode.py
+      - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
+      - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
 
 - label: ":nvidia: (L4) Quantization Kernels Shard %N"
   key: kernels-quantization-test
@@ -640,31 +685,6 @@ steps:
     - pytest -v -s kernels/moe/test_moe_layer.py
     - pytest -v -s kernels/moe/test_deepep_v2_moe.py
 
-- label: ":amd: (MI355) MLA Kernels"
-  key: amd-mla-kernels-mi355
-  dind: false
-  device: mi355_1
-  num_devices: 1
-  timeout_in_minutes: 25
-  working_dir: /vllm-workspace/
-  depends_on:
-  - image-build-amd
-  source_file_dependencies:
-  - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
-  - vllm/v1/attention/selector.py
-  - vllm/platforms/rocm.py
-  - vllm/_aiter_ops.py
-  - tests/kernels/attention/test_rocm_aiter_mla_decode.py
-  - tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
-  - tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
-  - tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
-  commands:
-  - amd-smi
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_fp8_support.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_op_registration.py
-
 - label: ":amd: (MI355) FP8 MoE Kernels"
   key: amd-fp8-moe-kernels-mi355
   dind: false
@@ -720,9 +740,7 @@ steps:
   - vllm/_aiter_ops.py
   - examples/basic/offline_inference/chat.py
   - tests/kernels/attention/test_attention_selector.py
-  - tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
   commands:
   - amd-smi
   - python3 examples/basic/offline_inference/chat.py --attention-backend TRITON_ATTN
   - pytest -v -s tests/kernels/attention/test_attention_selector.py
-  - pytest -v -s tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py
diff --git a/.buildkite/test_areas/models_basic.yaml b/.buildkite/test_areas/models_basic.yaml
index 6702eb6dd3..930c2dd590 100644
--- a/.buildkite/test_areas/models_basic.yaml
+++ b/.buildkite/test_areas/models_basic.yaml
@@ -119,6 +119,17 @@ steps:
   - tests/models/glm5next/
   commands:
     - pytest -v -s models/glm5next
+  mirror:
+    amd:
+      label: ":amd: (MI355) GLM5Next"
+      dind: false
+      device: mi355_1
+      timeout_in_minutes: 35
+      working_dir: "/vllm-workspace/tests"
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/platforms/rocm.py
 
 - label: ":nvidia: (H200) Qwen4 Exp"
   key: qwen4-exp-unit-tests
diff --git a/.buildkite/test_areas/models_language.yaml b/.buildkite/test_areas/models_language.yaml
index 3f80dfdfa3..c8473e7f0b 100644
--- a/.buildkite/test_areas/models_language.yaml
+++ b/.buildkite/test_areas/models_language.yaml
@@ -91,7 +91,7 @@ steps:
       commands:
       - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
       - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-      - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+      - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
 
 # Granite 4 hybrid generation is sensitive to hardware-specific Triton SSD
 # autotuning (https://github.com/vllm-project/vllm/issues/25194). Keep this one
@@ -111,6 +111,27 @@ steps:
     - rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
     - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
     - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
+  mirror:
+    amd:
+      label: ":amd: (MI355) Granite Language Model Compatibility"
+      dind: false
+      device: mi355_1
+      num_devices: 1
+      timeout_in_minutes: 80
+      working_dir: /vllm-workspace/tests
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/
+      - "!vllm/distributed/kv_transfer/"
+      - tests/models/language/generation
+      commands:
+      - export GIT_CONFIG_COUNT=1
+      - export GIT_CONFIG_KEY_0=http.version
+      - export GIT_CONFIG_VALUE_0=HTTP/1.1
+      - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
+      - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
+      - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
 
 - label: ":nvidia: (H200 MIG 35GB) Language Models (Extended Generation)"
   device: h200_35gb
diff --git a/.buildkite/test_areas/pytorch.yaml b/.buildkite/test_areas/pytorch.yaml
index 459f7faeb4..8204fb9670 100644
--- a/.buildkite/test_areas/pytorch.yaml
+++ b/.buildkite/test_areas/pytorch.yaml
@@ -87,6 +87,20 @@ steps:
     - tests/compile/h100/
   commands:
   - "find compile/h100/ -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+  mirror:
+    amd:
+      label: ":amd: (MI355) PyTorch Compilation H100 Cases"
+      dind: false
+      device: mi355_1
+      timeout_in_minutes: 45
+      working_dir: "/vllm-workspace/tests"
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - tests/utils.py
+      - csrc/
+      commands:
+      - pytest -v -s compile/h100/test_startup.py
 
 - label: ":nvidia: (L4) PyTorch Compilation Passes"
   key: pytorch-compilation-passes-unit-tests
diff --git a/.buildkite/test_areas/samplers.yaml b/.buildkite/test_areas/samplers.yaml
index e1bc4c08d8..084583c8a7 100644
--- a/.buildkite/test_areas/samplers.yaml
+++ b/.buildkite/test_areas/samplers.yaml
@@ -35,7 +35,7 @@ steps:
       - vllm/_aiter_ops.py
       - vllm/platforms/rocm.py
       commands:
-      - pytest -v -s samplers
+      - pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
 
 # Qwen2-Audio-7B leaves no room for KV cache on a 16 GiB slice.
 - label: ":nvidia: (H200 MIG 35GB) Samplers Multimodal Beam Search"
@@ -51,3 +51,24 @@ steps:
   commands:
     - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
+  mirror:
+    amd:
+      label: ":amd: (MI355) Samplers Multimodal Beam Search"
+      dind: false
+      device: mi355_1
+      num_devices: 1
+      timeout_in_minutes: 45
+      working_dir: /vllm-workspace/tests
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/model_executor/layers
+      - vllm/sampling_metadata.py
+      - vllm/v1/sample/
+      - vllm/entrypoints/generate/beam_search/
+      - tests/samplers
+      - tests/conftest.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
+      commands:
+      - pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
diff --git a/tests/compile/h100/test_startup.py b/tests/compile/h100/test_startup.py
index e57cf85b4a..a56cf15e5c 100644
--- a/tests/compile/h100/test_startup.py
+++ b/tests/compile/h100/test_startup.py
@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Cold start and warm start tests for vLLM-compile.
 
-Cold start runs in a forked child (must fork before CUDA init) which
+Cold start runs in a separate child process which
 populates on-disk caches and asserts cold-start counters.  Warm start
 then runs in the parent with clean in-memory state but populated caches.
 """
@@ -18,7 +18,7 @@ from vllm.compilation.counter import compilation_counter
 from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode, PassConfig
 from vllm.utils.torch_utils import is_torch_equal_or_newer
 
-from ...utils import fork_new_process_for_each_test
+from ...utils import create_new_process_for_each_test, requires_spawn_multiprocessing
 
 MODEL = "microsoft/Phi-tiny-MoE-instruct"
 
@@ -54,17 +54,21 @@ def _cold_start(vllm_runner):
     assert counters["aot_autograd"]["autograd_cache_hit"] == 0
 
 
-@fork_new_process_for_each_test
+@create_new_process_for_each_test()
 @pytest.mark.parametrize("mega_aot_artifact", ["0", "1"])
 def test_moe_startup(monkeypatch, vllm_runner, fresh_vllm_cache, mega_aot_artifact):
     monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
     monkeypatch.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", mega_aot_artifact)
     monkeypatch.setenv("VLLM_DEEP_GEMM_WARMUP", "skip")
 
-    # Cold start in a forked child (must fork before CUDA init).
+    # ROCm and XPU cannot fork safely after device discovery.
     # This model has 32 identical transformer layers which produce
     # 33 subgraphs after splitting on attention — only 3 are unique.
-    ctx = mp.get_context("fork")
+    ctx = (
+        mp.get_context("spawn")
+        if requires_spawn_multiprocessing()
+        else mp.get_context("fork")
+    )
     p = ctx.Process(target=_cold_start, args=(vllm_runner,))
     p.start()
     p.join()
@@ -223,13 +227,17 @@ def _cold_start_model(vllm_runner, spec: ModelStartupSpec):
 
 
 @pytest.mark.parametrize("spec", MODEL_SPECS)
-@fork_new_process_for_each_test
+@create_new_process_for_each_test()
 def test_model_startup(monkeypatch, vllm_runner, fresh_vllm_cache, spec):
     monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
     monkeypatch.setenv("VLLM_DEEP_GEMM_WARMUP", "skip")
 
-    # Cold start in a forked child (must fork before CUDA init).
-    ctx = mp.get_context("fork")
+    # ROCm and XPU cannot fork safely after device discovery.
+    ctx = (
+        mp.get_context("spawn")
+        if requires_spawn_multiprocessing()
+        else mp.get_context("fork")
+    )
     p = ctx.Process(target=_cold_start_model, args=(vllm_runner, spec))
     p.start()
     p.join()
diff --git a/tests/kernels/attention/test_triton_unified_attention.py b/tests/kernels/attention/test_triton_unified_attention.py
index 772fee137a..c3a2919fac 100644
--- a/tests/kernels/attention/test_triton_unified_attention.py
+++ b/tests/kernels/attention/test_triton_unified_attention.py
@@ -135,6 +135,7 @@ def ref_paged_attn(
     num_seqs = len(query_lens)
     block_tables = block_tables.cpu().numpy()
     _, block_size, num_kv_heads, head_size = key_cache.shape
+    head_size_v = value_cache.shape[-1]
 
     outputs: list[torch.Tensor] = []
     start_idx = 0
@@ -149,7 +150,7 @@ def ref_paged_attn(
 
         k = key_cache[block_indices].view(-1, num_kv_heads, head_size)
         k = k[:kv_len]
-        v = value_cache[block_indices].view(-1, num_kv_heads, head_size)
+        v = value_cache[block_indices].view(-1, num_kv_heads, head_size_v)
         v = v[:kv_len]
 
         if q.shape[1] != k.shape[1]:
diff --git a/tests/kernels/attention/test_triton_unified_attention_diffkv.py b/tests/kernels/attention/test_triton_unified_attention_diffkv.py
index 04b7755853..ea90b60a75 100644
--- a/tests/kernels/attention/test_triton_unified_attention_diffkv.py
+++ b/tests/kernels/attention/test_triton_unified_attention_diffkv.py
@@ -5,6 +5,7 @@
 import pytest
 import torch
 
+from tests.kernels.attention.test_triton_unified_attention import ref_paged_attn
 from vllm.platforms import current_platform
 from vllm.utils.math_utils import next_power_of_2
 from vllm.utils.torch_utils import (
@@ -81,7 +82,7 @@ def _alloc_segm_buffers(seq_threshold_3D: int, num_query_heads: int, head_size_v
 @pytest.mark.parametrize("dtype", DTYPES)
 @pytest.mark.parametrize("seq_threshold_3D", SEQ_THRESHOLD_3D_VALUES)
 @torch.inference_mode()
-def test_triton_unified_attn_diffkv_vs_fa(
+def test_triton_unified_attn_diffkv_vs_reference(
     seq_lens: list[tuple[int, int]],
     num_heads: tuple[int, int],
     head_sizes: tuple[int, int],
@@ -93,12 +94,13 @@ def test_triton_unified_attn_diffkv_vs_fa(
 ) -> None:
     head_size_qk, head_size_v = head_sizes
 
-    # DiffKV requires FA3 (Hopper) / FA4 (Blackwell) as the reference.
-    fa_version = get_flash_attn_version(head_size=head_size_qk, head_size_v=head_size_v)
-    if not is_flash_attn_varlen_func_available() or fa_version not in (3, 4):
-        pytest.skip(f"FA DiffKV needs FA3/FA4 (got version {fa_version}).")
-
-    from vllm.v1.attention.backends.fa_utils import flash_attn_varlen_func
+    # Keep the FA3/FA4 comparison on NVIDIA; ROCm uses the PyTorch oracle.
+    if not current_platform.is_rocm():
+        fa_version = get_flash_attn_version(
+            head_size=head_size_qk, head_size_v=head_size_v
+        )
+        if not is_flash_attn_varlen_func_available() or fa_version not in (3, 4):
+            pytest.skip(f"FA DiffKV needs FA3/FA4 (got version {fa_version}).")
 
     torch.set_default_device(DEVICE_TYPE)
     set_random_seed(0)
@@ -135,28 +137,44 @@ def test_triton_unified_attn_diffkv_vs_fa(
         0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
     )
 
-    # ---- FlashAttention DiffKV (ground truth) ---------------------------
-    # Mirror the backend: fix degenerate strides on size-1 dims so FA's
-    # TMA path sees ≥16-byte-aligned strides (matters for num_kv_heads==1).
-    fa_k = canonicalize_singleton_dim_strides(key_cache)
-    fa_v = canonicalize_singleton_dim_strides(value_cache)
-    fa_out = torch.empty(sum(query_lens), num_query_heads, head_size_v, dtype=dtype)
-    flash_attn_varlen_func(
-        q=query,
-        k=fa_k,
-        v=fa_v,
-        out=fa_out,
-        cu_seqlens_q=cu_query_lens,
-        max_seqlen_q=max_query_len,
-        seqused_k=kv_lens_t,
-        max_seqlen_k=max_kv_len,
-        softmax_scale=scale,
-        causal=True,
-        window_size=list(window_size),
-        block_table=block_tables,
-        softcap=soft_cap if soft_cap is not None else 0,
-        fa_version=fa_version,
-    )
+    if current_platform.is_rocm():
+        # FP32 also keeps the helper's in-place scaling off the kernel input.
+        ref_out = ref_paged_attn(
+            query.float(),
+            key_cache.float(),
+            value_cache.float(),
+            query_lens,
+            kv_lens,
+            block_tables,
+            scale,
+            sliding_window=sliding_window,
+            soft_cap=soft_cap,
+        ).to(dtype)
+    else:
+        from vllm.v1.attention.backends.fa_utils import flash_attn_varlen_func
+
+        # FA's TMA path needs aligned singleton strides (num_kv_heads == 1).
+        fa_k = canonicalize_singleton_dim_strides(key_cache)
+        fa_v = canonicalize_singleton_dim_strides(value_cache)
+        ref_out = torch.empty(
+            sum(query_lens), num_query_heads, head_size_v, dtype=dtype
+        )
+        flash_attn_varlen_func(
+            q=query,
+            k=fa_k,
+            v=fa_v,
+            out=ref_out,
+            cu_seqlens_q=cu_query_lens,
+            max_seqlen_q=max_query_len,
+            seqused_k=kv_lens_t,
+            max_seqlen_k=max_kv_len,
+            softmax_scale=scale,
+            causal=True,
+            window_size=list(window_size),
+            block_table=block_tables,
+            softcap=soft_cap if soft_cap is not None else 0,
+            fa_version=fa_version,
+        )
 
     # ---- Triton DiffKV --------------------------------------------------
     segm_output, segm_max, segm_expsum = _alloc_segm_buffers(
@@ -183,7 +201,4 @@ def test_triton_unified_attn_diffkv_vs_fa(
         softmax_segm_expsum=segm_expsum,
     )
 
-    (
-        torch.testing.assert_close(triton_out, fa_out, atol=2e-2, rtol=2e-2),
-        f"triton vs FA max abs diff: {torch.max(torch.abs(triton_out - fa_out))}",
-    )
+    torch.testing.assert_close(triton_out, ref_out, atol=2e-2, rtol=2e-2)
diff --git a/tests/models/glm5next/test_kda_recurrent.py b/tests/models/glm5next/test_kda_recurrent.py
index 68cf493b44..234a63fb73 100644
--- a/tests/models/glm5next/test_kda_recurrent.py
+++ b/tests/models/glm5next/test_kda_recurrent.py
@@ -4,18 +4,23 @@
 
 The decode path hands the kernel column slices of the merged ``q|k|v`` conv
 output and of the fused ``qkvbfg_a`` projection (beta), so q/k/v/beta are
-token-strided rather than contiguous. The kernel must read them in place,
-match a pure-PyTorch recurrence, and reject layouts it cannot address.
+token-strided rather than contiguous. Both backends must match a pure-PyTorch
+recurrence. CUDA reads these slices in place and rejects unsupported layouts;
+the ROCm wrapper makes contiguous copies before launching its kernel.
 """
 
 import pytest
 import torch
 
-from vllm.models.glm5next.nvidia.ops.third_party.kda import fused_recurrent_kda
 from vllm.platforms import current_platform
 
+if current_platform.is_rocm():
+    from vllm.models.glm5next.amd.ops.third_party.kda import fused_recurrent_kda
+else:
+    from vllm.models.glm5next.nvidia.ops.third_party.kda import fused_recurrent_kda
+
 pytestmark = pytest.mark.skipif(
-    not current_platform.is_cuda(), reason="CUDA-only Triton kernel"
+    not current_platform.is_cuda_alike(), reason="Requires CUDA or ROCm"
 )
 
 H, D = 16, 128
@@ -160,6 +165,10 @@ def test_fused_recurrent_kda_strided_inputs_bit_identical_to_contiguous(
     torch.testing.assert_close(state, state_ref, rtol=0, atol=0)
 
 
+@pytest.mark.skipif(
+    not current_platform.is_cuda(),
+    reason="Only the CUDA implementation requires zero-copy token-strided layouts",
+)
 @torch.inference_mode()
 def test_fused_recurrent_kda_rejects_unaddressable_layouts():
     """Layouts the token-stride addressing cannot express must fail loudly