[Bugfix][ROCm] AMD-Quark mixed-precision DeepSeek-V4.1 support (#57071)

Signed-off-by: Xiao Yu <xiao.yu.dc@outlook.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
xiao-llm
2026-09-25 20:51:18 -05:00
committed by GitHub
co-authored by Cursor
parent b6d2b4fc25
commit 5840d95284
11 changed files with 645 additions and 100 deletions
+79
View File
@@ -228,6 +228,85 @@ def test_deepseek_v41_vl_mapper_routes_linear_scales(
]
def test_deepseek_v41_vl_exposes_quant_mappings_on_class():
"""``configure_quant_config`` reads both mappings off the class, before
``__init__`` builds the instance mapper. Without them a Quark config's
per-layer keys never match this wrapper's ``language_model.``-rooted
prefixes and every attention shard falls back to the global spec."""
from vllm.models.deepseek_v41.amd.vl_model import DeepseekV41ForCausalLM
assert "fused_wqa_wkv" in DeepseekV41ForCausalLM.packed_modules_mapping
mapper = DeepseekV41ForCausalLM.hf_to_vllm_mapper.get_rename_mapper()
assert mapper.apply_list(["layers.0.attn.wq_a"]) == [
"language_model.model.layers.0.attn.wq_a"
]
def test_deepseek_v41_engram_scale_accepts_quark_name():
"""Quark exports name the engram scale ``embed.weight_scale``; the
``\\.scale$`` rules only match a literal ``.scale`` suffix, so without an
explicit rule the tensor is never routed and loading fails."""
from vllm.models.deepseek_v41.amd import model as model_module
mapper = model_module._make_deepseek_v4_weights_mapper("fp4", "weight_scale")
weight = torch.empty(0)
mapped = [
name
for name, _ in mapper.apply(
[
("layers.1.engram.embed.weight_scale", weight),
("layers.1.engram.embed.scale", weight),
]
)
]
assert mapped == [
"model.layers.1.engram.embed_tokens.weight_scale_inv",
"model.layers.1.engram.embed_tokens.weight_scale_inv",
]
def test_deepseek_v41_declines_quark_configs():
"""``from_config`` rewrites a Quark config into a single global FP8 scheme,
which would discard the per-layer specs of a mixed-precision export (MXFP4
experts + 2-D block MXFP8 attention in DeepSeek-V4.1-Flash). QuarkConfig
knows how to dispatch each scheme, so it must handle every Quark export."""
from vllm.models.deepseek_v41.quant_config import DeepseekV4FP8Config
hf_config = SimpleNamespace(model_type="deepseek_v41")
mxfp4_global = {
"global_quant_config": {
"weight": {"dtype": "fp4", "qscheme": "per_group", "group_size": 32}
},
"quant_method": "quark",
}
assert (
DeepseekV4FP8Config.override_quantization_method(
mxfp4_global, None, hf_config=hf_config
)
is None
)
mixed = {
**mxfp4_global,
"layer_quant_config": {
"layers.0.attn.wkv": {
"weight": {
"dtype": "fp8_e4m3",
"qscheme": "per_block",
"block_size": [32, 32],
}
}
},
}
assert (
DeepseekV4FP8Config.override_quantization_method(
mixed, None, hf_config=hf_config
)
is None
)
@pytest.mark.skipif(not current_platform.is_cuda(), reason="DeepGEMM requires CUDA")
@pytest.mark.parametrize("scale_dtype", [torch.uint8, torch.float8_e8m0fnu])
@pytest.mark.parametrize(
+282
View File
@@ -909,6 +909,184 @@ def test_quant_method_dispatch_target(case):
assert method_cls is (QuarkLinearMethod if is_linear else case.dispatch_cls)
def test_quant_method_dispatch_mxfp8_2d_block(default_vllm_config):
"""Shape B (DeepSeek-V4.1): 32x32 per-block e8m0 routes to QuarkOCP_MX.
Verifies that get_quant_method and get_quant_method_target agree (the
short-circuit that previously caused them to diverge has been removed).
"""
from vllm.model_executor.layers.quantization.quark.schemes import QuarkOCP_MX
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Static
default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16)
mxfp8_spec = {
"weight": {
"dtype": "fp8_e4m3",
"qscheme": "per_block",
"block_size": [32, 32],
"symmetric": True,
"is_dynamic": False,
"scale_type": "float8_e8m0fnu",
},
"input_tensors": {
"dtype": "fp8_e4m3",
"qscheme": "per_group",
"group_size": 32,
"symmetric": True,
"is_dynamic": True,
},
}
config = QuarkConfig(
{
"global_quant_config": {
"weight": {
"dtype": "fp4",
"qscheme": "per_group",
"group_size": 32,
"scale_format": "e8m0",
"is_dynamic": False,
},
"input_tensors": {
"dtype": "fp4",
"qscheme": "per_group",
"group_size": 32,
"scale_format": "e8m0",
"is_dynamic": True,
},
},
"layer_type_quant_config": {},
"layer_quant_config": {"layers.0.attn.wkv": mxfp8_spec},
"exclude": [],
}
)
class TestLinear(LinearBase):
def __init__(self):
torch.nn.Module.__init__(self)
# get_quant_method_target now routes MXFP8 through the matcher chain.
wk, ak, mcls = config.get_quant_method_target("layers.0.attn.wkv", LinearBase)
assert wk == kMxfp8Static
assert mcls is QuarkLinearMethod
linear = TestLinear()
method = config.get_quant_method(linear, "layers.0.attn.wkv")
assert isinstance(method, QuarkLinearMethod)
assert isinstance(linear.scheme, QuarkOCP_MX)
assert linear.scheme.weight_quant_key == kMxfp8Static
assert linear.scheme.scale_block_rows == 32
# Experts still fall through to the global MXFP4 spec.
assert (
config.get_quant_method_target("layers.0.ffn.experts", RoutedExperts)[2]
is not QuarkLinearMethod
)
# 128x128 per-block FP8 (DeepSeek V4) keeps its existing Quark scheme.
v4_spec = {
"weight": {**mxfp8_spec["weight"], "block_size": [128, 128]},
"input_tensors": {**mxfp8_spec["input_tensors"], "group_size": 128},
}
v4_config = _make_qtensor_config(v4_spec["weight"], v4_spec["input_tensors"])
assert v4_config.get_quant_method_target("linear", LinearBase)[2] is (
QuarkLinearMethod
)
linear = TestLinear()
assert isinstance(v4_config.get_quant_method(linear, "linear"), QuarkLinearMethod)
assert isinstance(linear.scheme, QuarkW8A8Fp8PerBlock)
def test_quant_method_dispatch_mxfp8_canonical(default_vllm_config):
"""Shape A (canonical MX): per_group/group_size=32/e8m0 FP8 routes
to QuarkOCP_MX with scale_block_rows == 1.
This covers the canonical MXFP8 spelling (no 2-D block). Validated by
synthetic dispatch test only — no canonical-MXFP8 checkpoint on hand.
"""
from vllm.model_executor.layers.quantization.quark.schemes import QuarkOCP_MX
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Static
default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16)
mxfp8_canonical_spec = {
"weight": {
"dtype": "fp8_e4m3",
"qscheme": "per_group",
"group_size": 32,
"scale_format": "e8m0",
"is_dynamic": False,
},
"input_tensors": {
"dtype": "fp8_e4m3",
"qscheme": "per_group",
"group_size": 32,
"scale_format": "e8m0",
"is_dynamic": True,
},
}
config = QuarkConfig(
{
"global_quant_config": mxfp8_canonical_spec,
"layer_type_quant_config": {},
"layer_quant_config": {},
"exclude": [],
}
)
wk, ak, mcls = config.get_quant_method_target("linear", LinearBase)
assert wk == kMxfp8Static
assert mcls is QuarkLinearMethod
class TestLinear(LinearBase):
def __init__(self):
torch.nn.Module.__init__(self)
linear = TestLinear()
method = config.get_quant_method(linear, "linear")
assert isinstance(method, QuarkLinearMethod)
assert isinstance(linear.scheme, QuarkOCP_MX)
assert linear.scheme.weight_quant_key == kMxfp8Static
assert linear.scheme.scale_block_rows == 1
def test_quant_method_dispatch_mxfp8_moe_raises(default_vllm_config):
"""MXFP8 in a MoE config raises ValueError — experts are unsupported."""
from vllm.model_executor.layers.quantization.quark.quark_moe import (
QuarkOCP_MX_MoEMethod,
)
default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16)
mxfp8_spec = {
"weight": {
"dtype": "fp8_e4m3",
"qscheme": "per_group",
"group_size": 32,
"scale_format": "e8m0",
"is_dynamic": False,
},
"input_tensors": {
"dtype": "fp8_e4m3",
"qscheme": "per_group",
"group_size": 32,
"scale_format": "e8m0",
"is_dynamic": True,
},
}
config = QuarkConfig(
{
"global_quant_config": mxfp8_spec,
"layer_type_quant_config": {},
"layer_quant_config": {},
"exclude": [],
}
)
wk, ak, mcls = config.get_quant_method_target("experts", RoutedExperts)
assert mcls is QuarkOCP_MX_MoEMethod
# The OCP MX MoE constructor should fail loudly for MXFP8.
fake_moe_config = MagicMock()
with pytest.raises(ValueError, match="MXFP8 experts are not supported"):
QuarkOCP_MX_MoEMethod(fake_moe_config, wk, ak)
@pytest.mark.parametrize(
("weight", "input_tensors"),
[
@@ -2313,3 +2491,107 @@ def test_quark_int4_canonicalizes_pack_for_kernel_layout(pack_method, symmetric)
)
assert torch.equal(actual, expected)
# ---------------------------------------------------------------------------
# override_quantization_method: Quark always delegates to QuarkConfig
# ---------------------------------------------------------------------------
_QUARK_MXFP4_CFG = {
"quant_method": "quark",
"global_quant_config": {
"weight": {"dtype": "fp4", "qscheme": "per_group", "group_size": 32},
},
}
_QUARK_MXFP8_CFG = {
"quant_method": "quark",
"global_quant_config": {
"weight": {
"dtype": "fp8_e4m3",
"qscheme": "per_block",
"block_size": [32, 32],
"scale_type": "float8_e8m0fnu",
"symmetric": True,
},
},
}
_QUARK_MIXED_CFG = {
"quant_method": "quark",
"global_quant_config": {
"weight": {"dtype": "fp4", "qscheme": "per_group", "group_size": 32},
},
"layer_quant_config": {"layers.0.attn": {"weight": {"dtype": "fp8_e4m3"}}},
}
_FP8_CFG = {"quant_method": "fp8"}
_DSV4_FP8_CFG = {"quant_method": "deepseek_v4_fp8"}
@pytest.mark.parametrize(
"quant_config_cls",
[
"deepseek_v41",
"deepseek_v4",
],
)
@pytest.mark.parametrize(
"hf_quant_cfg, expect_none",
[
(_QUARK_MXFP4_CFG, True),
(_QUARK_MXFP8_CFG, True),
(_QUARK_MIXED_CFG, True),
(_FP8_CFG, False),
(_DSV4_FP8_CFG, False),
],
ids=[
"quark-mxfp4",
"quark-mxfp8",
"quark-mixed",
"native-fp8",
"deepseek-v4-fp8",
],
)
def test_quark_override_delegates_to_quark_config(
quant_config_cls, hf_quant_cfg, expect_none
):
"""Any quant_method=='quark' must return None (delegate to QuarkConfig).
Native FP8 configs should still be claimed by the model-specific config.
"""
import importlib
if quant_config_cls == "deepseek_v41":
# Module was renamed from deepseek_v4_1 to deepseek_v41 upstream.
for mod_name in (
"vllm.models.deepseek_v41.quant_config",
"vllm.models.deepseek_v4_1.quant_config",
):
try:
mod = importlib.import_module(mod_name)
break
except ModuleNotFoundError:
continue
else:
pytest.skip("deepseek_v41 quant_config not found")
model_type = "deepseek_v41"
else:
mod = importlib.import_module("vllm.models.deepseek_v4.quant_config")
model_type = "deepseek_v4"
config_cls = mod.DeepseekV4FP8Config
hf_config = SimpleNamespace(model_type=model_type)
result = config_cls.override_quantization_method(
hf_quant_cfg, None, hf_config=hf_config
)
if expect_none:
assert result is None, (
f"Quark config should delegate to QuarkConfig (return None), got {result!r}"
)
else:
assert result == "deepseek_v4_fp8", (
f"Native FP8 should be claimed by DeepseekV4FP8Config, got {result!r}"
)