[Test] [AMD] diffusion: Make multimodal_gen unit tests ROCm-aware (#44599)

This commit is contained in:
Michael
2026-10-20 14:24:37 +08:00
committed by GitHub
parent 4b83f215f4
commit 407bfc94f9
7 changed files with 48 additions and 6 deletions
@@ -2,6 +2,7 @@
"""CPU-safe FLUX.3 tests for the rowwise-FP8 and ModelOpt NVFP4 DiT paths."""
import contextlib
import importlib.util
import json
import tempfile
import unittest
@@ -9,6 +10,7 @@ from pathlib import Path
from types import SimpleNamespace
from unittest import mock
import pytest
import torch
from safetensors.torch import save_file
@@ -38,6 +40,14 @@ from sglang.multimodal_gen.runtime.pipelines.flux3_image_weights import (
from sglang.multimodal_gen.runtime.platforms import current_platform
from sglang.multimodal_gen.runtime.platforms.interface import DeviceCapability
# Building the DiT constructs attention layers, which read the SP group size.
pytestmark = pytest.mark.usefixtures("single_process_model_parallel")
_requires_flashinfer = unittest.skipUnless(
importlib.util.find_spec("flashinfer"),
"the PrimsTS weight pack ships with flashinfer (CUDA-only)",
)
_LEAVES = ("to_qkv", "mlp_in", "attn_out", "mlp_out")
_ALL_ROLES = {
"to_qkv": Flux3PrimsTsRole.QKV_QKNORM_ROPE,
@@ -207,6 +217,7 @@ def _roles(block):
class TestFlux3Fp8Construction(unittest.TestCase):
@_requires_flashinfer
def test_process_weights_prepares_prims_ts_weights(self):
from sglang.kernels.ops.diffusion import pack_flux3_mlp_weight
@@ -258,6 +269,7 @@ class TestFlux3Fp8Construction(unittest.TestCase):
self.assertEqual(block.to_qkv.output_size_per_partition, 768)
self.assertEqual(_roles(block), {**_ALL_ROLES, "to_qkv": None})
@_requires_flashinfer
def test_swiglu_layout_rejects_generic_apply(self):
layer = _tiny_native_model(_flux3_fp8_config()).single_blocks[0].mlp_in
layer.quant_method.process_weights_after_loading(layer)
@@ -270,6 +282,7 @@ class TestFlux3Fp8Construction(unittest.TestCase):
self.assertIsInstance(block.to_qkv.quant_method, Flux3Fp8LinearMethod)
self.assertIsInstance(block.mlp_in.quant_method, UnquantizedLinearMethod)
@_requires_flashinfer
def test_fused_mlp_weight_pack_reverses_each_gate_value_pair(self):
from sglang.kernels.ops.diffusion import pack_flux3_mlp_weight
@@ -4,7 +4,11 @@
import unittest
from unittest import mock
from flashinfer.prims_ts import gemm as prims_ts_gemm
import pytest
prims_ts_gemm = pytest.importorskip(
"flashinfer.prims_ts.gemm", reason="PrimsTS ships with flashinfer (CUDA-only)"
)
from sglang.kernels.ops.diffusion.gemm import flux3_prims_ts_config as config_module
@@ -36,7 +36,10 @@ from sglang.multimodal_gen.runtime.models.vaes.kandinsky6_audio import (
Kandinsky6AudioVAE,
)
from sglang.multimodal_gen.runtime.models.vaes.kandinsky6_sr_vae import Kandinsky6SRVAE
from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
from sglang.multimodal_gen.runtime.platforms import (
AttentionBackendEnum,
current_platform,
)
from sglang.multimodal_gen.runtime.server_args import get_global_server_args
from sglang.multimodal_gen.runtime.utils import precision
from sglang.multimodal_gen.test.unit.kandinsky6_sr_tiny_components import (
@@ -219,6 +222,11 @@ def test_layerwise_groups_preserve_repeated_forwards(component, resident_layers)
assert set(groups) <= set(model.layer_names)
assert model.layerwise_offload_dit_group_enabled == component.endswith("dit")
state = {name: value.cpu().clone() for name, value in model.state_dict().items()}
tolerance = dict(rtol=0, atol=0)
if component == "audio-vae" and current_platform.is_rocm():
# On MI300 the offloaded fp32 audio VAE differs from the resident run by
# ~1e-6 (CUDA is bit-exact); the weights are still compared exactly below.
tolerance = dict(rtol=1e-4, atol=1e-5)
with (
torch.inference_mode(),
set_forward_context(current_timestep=0, attn_metadata=None),
@@ -241,12 +249,12 @@ def test_layerwise_groups_preserve_repeated_forwards(component, resident_layers)
} == set(groups)
for _ in range(2):
model.prepare_for_next_req()
torch.testing.assert_close(forward(), expected, rtol=0, atol=0)
torch.testing.assert_close(forward(), expected, **tolerance)
for manager in model.layerwise_offload_managers:
assert manager._last_forwarded_layer == manager.num_layers - 1
manager.release_all()
model.disable_offload()
torch.testing.assert_close(forward(), expected, rtol=0, atol=0)
torch.testing.assert_close(forward(), expected, **tolerance)
for name, value in model.state_dict().items():
torch.testing.assert_close(value.cpu(), state[name], rtol=0, atol=0)
@@ -128,6 +128,11 @@ def _worker(mode):
dist.destroy_process_group()
@pytest.mark.skipif(
torch.version.hip is not None,
reason="ROCm builds query GPU 0 while importing sglang, which fails with the "
"GPUs hidden for these CPU workers",
)
@pytest.mark.parametrize("mode", ["sp", "cfg"])
def test_two_rank_moss_image_matches_serial(mode):
result = subprocess.run(
@@ -101,10 +101,15 @@ def test_native_encoder_registry_and_checkpoint(small_text_encoder):
@pytest.mark.parametrize("backend", [None, "torch_sdpa", "fa"])
def test_text_attention_default_and_explicit_selection(small_text_encoder, backend):
from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
from sglang.multimodal_gen.runtime.platforms import (
AttentionBackendEnum,
current_platform,
)
if backend == "fa" and not torch.cuda.is_available():
pytest.skip("FA selection requires CUDA")
if backend == "fa" and current_platform.is_rocm():
pytest.skip("the platform currently resolves FA to SDPA on ROCm")
get_global_server_args().attention_backend = backend
config = Qwen3_5EncoderConfig()
config.update_model_arch(small_text_encoder.config.to_dict())
@@ -25,6 +25,11 @@ class TestResidualGateAdd2Cuda(unittest.TestCase):
# Raises instead of falling back when the predicate rejects.
got = residual_gate_add2_cuda(residual, a, b, gate)
want = torch.addcmul(residual, gate, a + b)
if torch.version.hip:
# On MI300 the fp16 scalar path is not bit-identical to
# addcmul; hold it to the dtype's rounding tolerance.
torch.testing.assert_close(got, want)
continue
self.assertTrue(torch.equal(got, want), (dtype, shape, gate.shape))
expanded_gate = full_gate[:1].expand_as(residual)
self.assertFalse(
@@ -174,7 +174,9 @@ class TestRowwiseFp8Config(unittest.TestCase):
def _fp8_capable():
if not torch.cuda.is_available():
# ROCm only takes FP8 weights through the loader, which converts them to
# the platform's FP8 layout; this test feeds raw OCP e4m3fn weights.
if not torch.cuda.is_available() or torch.version.hip:
return False
major, minor = torch.cuda.get_device_capability()
return major >= 9 or (major == 8 and minor >= 9)