rocm.py 11.9 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import os
4
from functools import cache, lru_cache, wraps
5
from typing import TYPE_CHECKING, Dict, List, Optional
6
7
8

import torch

9
import vllm.envs as envs
10
11
from vllm.logger import init_logger

12
from .interface import DeviceCapability, Platform, PlatformEnum, _Backend
13

14
if TYPE_CHECKING:
15
    from vllm.config import ModelConfig, VllmConfig
16
else:
17
    ModelConfig = None
18
    VllmConfig = None
19

20
21
logger = init_logger(__name__)

22
try:
23
24
25
    from amdsmi import (AmdSmiException, amdsmi_get_gpu_asic_info,
                        amdsmi_get_processor_handles, amdsmi_init,
                        amdsmi_shut_down, amdsmi_topo_get_link_type)
26
27
28
except ImportError as e:
    logger.warning("Failed to import from amdsmi with %r", e)

29
30
31
32
33
34
try:
    import vllm._C  # noqa: F401
except ImportError as e:
    logger.warning("Failed to import from vllm._C with %r", e)

# import custom ops, trigger op registration
zhuwenwen's avatar
zhuwenwen committed
35
36
37
38
# try:
#     import vllm._rocm_C  # noqa: F401
# except ImportError as e:
#     logger.warning("Failed to import from vllm._rocm_C with %r", e)
39

40

41
42
43
44
45
# Models not supported by ROCm.
_ROCM_UNSUPPORTED_MODELS: List[str] = []

# Models partially supported by ROCm.
# Architecture -> Reason.
zhuwenwen's avatar
zhuwenwen committed
46
47
48
49
# _ROCM_SWA_REASON = ("Sliding window attention (SWA) is not yet supported in "
#                     "Triton flash attention. For half-precision SWA support, "
#                     "please use CK flash attention by setting "
#                     "`VLLM_USE_TRITON_FLASH_ATTN=0`")
50
_ROCM_PARTIALLY_SUPPORTED_MODELS: Dict[str, str] = {
zhuwenwen's avatar
zhuwenwen committed
51
52
53
54
55
56
    # "Qwen2ForCausalLM":
    # _ROCM_SWA_REASON,
    # "MistralForCausalLM":
    # _ROCM_SWA_REASON,
    # "MixtralForCausalLM":
    # _ROCM_SWA_REASON,
57
58
59
    "PaliGemmaForConditionalGeneration":
    ("ROCm flash attention does not yet "
     "fully support 32-bit precision on PaliGemma"),
zhuwenwen's avatar
zhuwenwen committed
60
61
62
63
    # "Phi3VForCausalLM":
    # ("ROCm Triton flash attention may run into compilation errors due to "
    #  "excessive use of shared memory. If this happens, disable Triton FA "
    #  "by setting `VLLM_USE_TRITON_FLASH_ATTN=0`")
64
65
}

66
# Prevent use of clashing `{CUDA/HIP}_VISIBLE_DEVICES``
67
68
69
70
71
72
# if "HIP_VISIBLE_DEVICES" in os.environ:
#     val = os.environ["HIP_VISIBLE_DEVICES"]
#     if cuda_val := os.environ.get("CUDA_VISIBLE_DEVICES", None):
#         assert val == cuda_val
#     else:
#         os.environ["CUDA_VISIBLE_DEVICES"] = val
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100

# AMDSMI utils
# Note that NVML is not affected by `{CUDA/HIP}_VISIBLE_DEVICES`,
# all the related functions work on real physical device ids.
# the major benefit of using AMDSMI is that it will not initialize CUDA


def with_amdsmi_context(fn):

    @wraps(fn)
    def wrapper(*args, **kwargs):
        amdsmi_init()
        try:
            return fn(*args, **kwargs)
        finally:
            amdsmi_shut_down()

    return wrapper


def device_id_to_physical_device_id(device_id: int) -> int:
    if "CUDA_VISIBLE_DEVICES" in os.environ:
        device_ids = os.environ["CUDA_VISIBLE_DEVICES"].split(",")
        physical_device_id = device_ids[device_id]
        return int(physical_device_id)
    else:
        return device_id

101

102
103
104
105
106
107
108
109
110
111
112
@cache
def use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
                                    block_size: int, gqa_ratio: int,
                                    max_seq_len: int,
                                    sliding_window: int) -> bool:

    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    ON_NAVI = "gfx1" in GPU_ARCH
    ON_MI250_MI300 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942"])

    # rocm custom page attention not support on navi (gfx1*)
113
114
    return (ON_MI250_MI300 and not ON_NAVI
            and (sliding_window == 0 or sliding_window == (-1, -1))
115
116
117
118
119
120
121
            and (qtype == torch.half or qtype == torch.bfloat16)
            and (head_size == 64 or head_size == 128)
            and (block_size == 16 or block_size == 32)
            and (gqa_ratio >= 1 and gqa_ratio <= 16) and max_seq_len <= 32768
            and envs.VLLM_ROCM_CUSTOM_PAGED_ATTN)


122
123
class RocmPlatform(Platform):
    _enum = PlatformEnum.ROCM
124
    device_name: str = "rocm"
125
    device_type: str = "cuda"
126
    dispatch_key: str = "CUDA"
127
    ray_device_key: str = "GPU"
128
129
    # rocm shares the same device control env var as CUDA
    device_control_env_var: str = "CUDA_VISIBLE_DEVICES"
130

131
132
    supported_quantization: list[str] = [
        "awq", "gptq", "fp8", "compressed_tensors", "compressed-tensors",
133
        "fbgemm_fp8", "gguf", "quark", "ptpc_fp8", "moe_wna16", "blockwise_int8"
134
    ]
135

136
    @classmethod
137
    def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
138
139
                             kv_cache_dtype, block_size, use_v1,
                             use_mla) -> str:
140
141
142
        if use_mla:
            logger.info("Using Triton MLA backend.")
            return "vllm.attention.backends.triton_mla.TritonMLABackend"
143
144
        selected_backend = (_Backend.ROCM_FLASH if selected_backend
                            == _Backend.FLASH_ATTN else selected_backend)
145
        if envs.VLLM_USE_V1:
146
147
148
            logger.info("Using Triton Attention backend on V1 engine.")
            return ("vllm.v1.attention.backends."
                    "triton_attn.TritonAttentionBackend")
149
150
151
152
153
154
        if selected_backend == _Backend.ROCM_FLASH:
            if not cls.has_device_capability(90):
                # not Instinct series GPUs.
                logger.info("flash_attn is not supported on NAVI GPUs.")
        else:
            logger.info("%s is not supported in AMD GPUs.", selected_backend)
155
156
        logger.info("Using ROCmFlashAttention backend.")
        return "vllm.attention.backends.rocm_flash_attn.ROCmFlashAttentionBackend"  # noqa: E501
157

158
    @classmethod
159
    @lru_cache(maxsize=8)
160
161
162
    def get_device_capability(cls,
                              device_id: int = 0
                              ) -> Optional[DeviceCapability]:
163
164
        major, minor = torch.cuda.get_device_capability(device_id)
        return DeviceCapability(major=major, minor=minor)
165

zhuwenwen's avatar
zhuwenwen committed
166
    @staticmethod
167
168
    @with_amdsmi_context
    def is_fully_connected(physical_device_ids: List[int]) -> bool:
zhuwenwen's avatar
zhuwenwen committed
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
        """
        Query if the set of gpus are fully connected by xgmi (1 hop)
        """
        handles = [
            amdsmi_get_processor_handles()[i] for i in physical_device_ids
        ]
        for i, handle in enumerate(handles):
            for j, peer_handle in enumerate(handles):
                if i < j:
                    try:
                        link_type = amdsmi_topo_get_link_type(
                            handle, peer_handle)
                        # type is 2 for XGMI
                        if link_type["hops"] != 1 or link_type["type"] != 2:
                            return False
                    except AmdSmiException as error:
                        logger.error("AMD 1 hop XGMI detection failed.",
                                     exc_info=error)
                        return False
        return True
189

190
    @classmethod
191
    @with_amdsmi_context
192
    @lru_cache(maxsize=8)
193
    def get_device_name(cls, device_id: int = 0) -> str:
194
195
        physical_device_id = device_id_to_physical_device_id(device_id)
        handle = amdsmi_get_processor_handles()[physical_device_id]
zhuwenwen's avatar
zhuwenwen committed
196
197
        # return amdsmi_get_gpu_asic_info(handle)["market_name"]
        return torch.cuda.get_device_name(device_id)
198

zhuwenwen's avatar
zhuwenwen committed
199
200
201
202
    @classmethod
    def get_device_total_memory(cls, device_id: int = 0) -> int:
        device_props = torch.cuda.get_device_properties(device_id)
        return device_props.total_memory
203

204
205
206
207
208
209
210
211
212
213
    @classmethod
    def is_async_output_supported(cls, enforce_eager: Optional[bool]) -> bool:
        if enforce_eager:
            logger.warning(
                "To see benefits of async output processing, enable CUDA "
                "graph. Since, enforce-eager is enabled, async output "
                "processor cannot be used")
            return False
        return True

214
215
    @classmethod
    def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
216
217
218
219
        cache_config = vllm_config.cache_config
        if cache_config and cache_config.block_size is None:
            cache_config.block_size = 16

220
221
222
223
        parallel_config = vllm_config.parallel_config
        scheduler_config = vllm_config.scheduler_config
        if parallel_config.worker_cls == "auto":
            if scheduler_config.is_multi_step:
224
225
226
                if envs.VLLM_USE_V1:
                    raise NotImplementedError(
                        "Multi-step scheduling is not supported (and not "
227
                        "needed) on vLLM V1. Please launch without "
228
229
230
231
                        "--num-scheduler-steps.")
                else:
                    parallel_config.worker_cls = \
                        "vllm.worker.multi_step_worker.MultiStepWorker"
232
            elif vllm_config.speculative_config:
233
234
                if envs.VLLM_USE_V1:
                    raise NotImplementedError(
235
                        "Speculative decoding is not yet supported on vLLM V1."
236
237
238
239
240
241
                    )
                else:
                    parallel_config.worker_cls = \
                        "vllm.spec_decode.spec_decode_worker.create_spec_worker"
                    parallel_config.sd_worker_cls = \
                        "vllm.worker.worker.Worker"
242
            else:
243
244
245
246
247
                if envs.VLLM_USE_V1:
                    parallel_config.worker_cls = \
                            "vllm.v1.worker.gpu_worker.Worker"
                else:
                    parallel_config.worker_cls = "vllm.worker.worker.Worker"
248

249
250
251
252
253
254
255
256
257
258
259
260
    @classmethod
    def verify_model_arch(cls, model_arch: str) -> None:
        if model_arch in _ROCM_UNSUPPORTED_MODELS:
            raise ValueError(f"Model architecture '{model_arch}' is not "
                             "supported by ROCm for now.")

        if model_arch in _ROCM_PARTIALLY_SUPPORTED_MODELS:
            msg = _ROCM_PARTIALLY_SUPPORTED_MODELS[model_arch]
            logger.warning(
                "Model architecture '%s' is partially "
                "supported by ROCm: %s", model_arch, msg)

261
262
263
264
265
266
    @classmethod
    def verify_quantization(cls, quant: str) -> None:
        super().verify_quantization(quant)
        if quant == "awq" and not envs.VLLM_USE_TRITON_AWQ:
            logger.warning(
                "Using AWQ quantization with ROCm, but VLLM_USE_TRITON_AWQ"
267
268
                " is not set, disabling VLLM_USE_TRITON_AWQ.")
            envs.VLLM_USE_TRITON_AWQ = False
269
270
271
272

    @classmethod
    def get_punica_wrapper(cls) -> str:
        return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU"
273
274
275
276
277
278

    @classmethod
    def get_current_memory_usage(cls,
                                 device: Optional[torch.types.Device] = None
                                 ) -> float:
        torch.cuda.reset_peak_memory_stats(device)
zhuwenwen's avatar
zhuwenwen committed
279
280
281
        # return torch.cuda.mem_get_info(device)[1] - torch.cuda.mem_get_info(
        #     device)[0]
        return torch.cuda.max_memory_allocated(device)
282
283
284
285

    @classmethod
    def get_device_communicator_cls(cls) -> str:
        return "vllm.distributed.device_communicators.cuda_communicator.CudaCommunicator"  # noqa
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302

    @classmethod
    def supports_fp8(cls) -> bool:
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
        return any(gfx in gcn_arch for gfx in ['gfx94', 'gfx95', 'gfx12'])

    @classmethod
    def is_fp8_fnuz(cls) -> bool:
        # only device 0 is checked, this assumes MI300 platforms are homogeneous
        return 'gfx94' in torch.cuda.get_device_properties(0).gcnArchName

    @classmethod
    def fp8_dtype(cls) -> torch.dtype:
        if cls.is_fp8_fnuz():
            return torch.float8_e4m3fnuz
        else:
            return torch.float8_e4m3fn
303
304
305
306
307

    @classmethod
    def supports_v1(cls, model_config: ModelConfig) -> bool:
        # V1 support on AMD gpus is experimental
        return True
308
309
310
311
312
313
314

    @classmethod
    def use_custom_allreduce(cls) -> bool:
        # We only enable custom allreduce for MI300 series
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
        supported_archs = ['gfx94']
        return any(gfx in gcn_arch for gfx in supported_archs)