rocm.py 20.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import os
5
from datetime import timedelta
6
from functools import cache, lru_cache, wraps
7
from typing import TYPE_CHECKING, Optional
8
9

import torch
10
11
from torch.distributed import PrefixStore, ProcessGroup
from torch.distributed.distributed_c10d import is_nccl_available
12

13
import vllm.envs as envs
14
from vllm.logger import init_logger
15
from vllm.utils import cuda_device_count_stateless
16

17
from .interface import DeviceCapability, Platform, PlatformEnum, _Backend
18

19
if TYPE_CHECKING:
20
    from vllm.config import ModelConfig, VllmConfig
21

22
23
logger = init_logger(__name__)

24
try:
25
26
27
    from amdsmi import (AmdSmiException, amdsmi_get_gpu_asic_info,
                        amdsmi_get_processor_handles, amdsmi_init,
                        amdsmi_shut_down, amdsmi_topo_get_link_type)
28
29
30
except ImportError as e:
    logger.warning("Failed to import from amdsmi with %r", e)

31
32
33
34
35
36
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
37
38
39
40
# try:
#     import vllm._rocm_C  # noqa: F401
# except ImportError as e:
#     logger.warning("Failed to import from vllm._rocm_C with %r", e)
41

42

43
# Models not supported by ROCm.
44
_ROCM_UNSUPPORTED_MODELS: list[str] = []
45
46
47

# Models partially supported by ROCm.
# Architecture -> Reason.
zhuwenwen's avatar
zhuwenwen committed
48
49
50
51
# _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`")
52
_ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {
zhuwenwen's avatar
zhuwenwen committed
53
54
55
56
57
58
    # "Qwen2ForCausalLM":
    # _ROCM_SWA_REASON,
    # "MistralForCausalLM":
    # _ROCM_SWA_REASON,
    # "MixtralForCausalLM":
    # _ROCM_SWA_REASON,
zhuwenwen's avatar
zhuwenwen committed
59
60
61
    # "PaliGemmaForConditionalGeneration":
    # ("ROCm flash attention does not yet "
    #  "fully support 32-bit precision on PaliGemma"),
zhuwenwen's avatar
zhuwenwen committed
62
63
64
65
    # "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`")
66
}
67
_ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = {
68
69
70
71
72
73
74
75
    "0x74a0": "AMD_Instinct_MI300A",
    "0x74a1": "AMD_Instinct_MI300X",
    "0x74b5": "AMD_Instinct_MI300X",  # MI300X VF
    "0x74a5": "AMD_Instinct_MI325X",
    "0x74b9": "AMD_Instinct_MI325X",  # MI325X VF
    "0x74a9": "AMD_Instinct_MI300X_HF",
    "0x74bd": "AMD_Instinct_MI300X_HF",
}
76

77
# Prevent use of clashing `{CUDA/HIP}_VISIBLE_DEVICES``
zhuwenwen's avatar
zhuwenwen committed
78
79
80
81
82
83
# 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
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103

# 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


104
105
106
107
108
109
110
111
112
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


113
114
115
116
@cache
def on_gfx1x() -> bool:
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    return any(arch in GPU_ARCH for arch in ["gfx11", "gfx12"])
117

118

119
@cache
120
def on_mi3xx() -> bool:
121
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
122
    return any(arch in GPU_ARCH for arch in ["gfx942", "gfx950"])
123
124


125
@cache
126
127
128
def on_gfx9() -> bool:
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    return any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
129
130


131
@cache
132
133
134
135
136
137
138
139
140
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,
        kv_cache_dtype: str,
        alibi_slopes: Optional[torch.Tensor] = None) -> bool:
141

142
143
    GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
    ON_GFX9 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
144
    ON_GFX11_GFX12 = any(arch in GPU_ARCH for arch in ["gfx11", "gfx12"])
145

146
147
    # custom paged attn always supported on V0. On V1, requires sliding window
    # disabled due to observed numerical discrepancy.
zhuwenwen's avatar
zhuwenwen committed
148
149
150
151
152
153
154
    # if ON_GFX9:
    #     return ((not envs.VLLM_USE_V1 or sliding_window == 0
    #              or sliding_window == (-1, -1))
    #             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)
zhuwenwen's avatar
zhuwenwen committed
155
156
    #             and max_seq_len <= 128 * 1024
    #             and (envs.VLLM_ROCM_CUSTOM_PAGED_ATTN)
zhuwenwen's avatar
zhuwenwen committed
157
158
159
160
161
162
163
164
165
    #             and not (envs.VLLM_ROCM_USE_AITER_PAGED_ATTN
    #                      and envs.VLLM_ROCM_USE_AITER))

    # else:
    #     return (ON_GFX11_GFX12 and (not envs.VLLM_USE_V1 or sliding_window == 0
    #                                 or sliding_window == (-1, -1))
    #             and (qtype == torch.half or qtype == torch.bfloat16)
    #             and head_size == 128 and block_size == 16
    #             and (gqa_ratio >= 3 and gqa_ratio <= 16)
zhuwenwen's avatar
zhuwenwen committed
166
    #             and max_seq_len <= 128 * 1024 and alibi_slopes is None
zhuwenwen's avatar
zhuwenwen committed
167
168
    #             and kv_cache_dtype == "auto"
    #             and envs.VLLM_ROCM_CUSTOM_PAGED_ATTN)
169
    return False
170
171


172
173
class RocmPlatform(Platform):
    _enum = PlatformEnum.ROCM
174
    device_name: str = "rocm"
175
    device_type: str = "cuda"
176
    dispatch_key: str = "CUDA"
177
    ray_device_key: str = "GPU"
178
    dist_backend: str = "nccl"
179
180
    # rocm shares the same device control env var as CUDA
    device_control_env_var: str = "CUDA_VISIBLE_DEVICES"
181

182
    supported_quantization: list[str] = [
183
        "awq", "gptq", "fp8", "compressed-tensors", "fbgemm_fp8", "gguf",
184
        "quark", "ptpc_fp8", "moe_wna16", "slimquant_w4a8","w8a8_int8","awq_marlin"
185
    ]
186

187
    @classmethod
188
    def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
189
190
                             kv_cache_dtype, block_size, use_v1,
                             use_mla) -> str:
191
        if use_mla:
zhuwenwen's avatar
zhuwenwen committed
192
193
194
195
196
            if selected_backend == _Backend.TRITON_MLA or block_size != 64:
                if use_v1:
                    logger.info_once("Using Triton MLA backend on V1 engine.")
                    return ("vllm.v1.attention.backends.mla."
                            "triton_mla.TritonMLABackend")
197

zhuwenwen's avatar
zhuwenwen committed
198
199
                else:
                    logger.info("Using Triton MLA backend.")
zhuwenwen's avatar
zhuwenwen committed
200
201
                    return "vllm.attention.backends.triton_mla.TritonMLABackend"   
            else:         
202
203
204
205
206
207
208
209
210
211
212
213
                if envs.VLLM_USE_FLASH_MLA:
                    from vllm.attention.backends.flashmla import (
                        is_flashmla_supported)
                    if not is_flashmla_supported()[0]:
                        logger.warning(
                            "FlashMLA backend is not supported due to %s",
                            is_flashmla_supported()[1])
                    elif block_size != 64:
                        logger.warning(
                            "FlashMLA backend is not supported for block size %d"
                            " (currently only supports block size 64).",
                            block_size)
zhuwenwen's avatar
zhuwenwen committed
214
                    else:
215
216
217
218
219
220
221
222
223
224
225
                        if use_v1:
                            logger.info_once(
                                "Using FlashMLA backend on V1 engine.")
                            return ("vllm.v1.attention.backends.mla."
                                    "flashmla.FlashMLABackend")
                        else:
                            logger.info("Using FlashMLA backend.")
                            return ("vllm.attention.backends."
                                    "flashmla.FlashMLABackend")
                else:
                    logger.info("Using Triton MLA backend (block size 64).")
zhuwenwen's avatar
zhuwenwen committed
226
                    return "vllm.attention.backends.triton_mla.TritonMLABackend"           
zhuwenwen's avatar
zhuwenwen committed
227
228
229
230
231
232
233
234
235
236
237

            # from vllm.attention.backends.rocm_aiter_mla import (
            #     is_aiter_mla_enabled)

            # if selected_backend is None:
            #     selected_backend = (_Backend.ROCM_AITER_MLA if
            #                         is_aiter_mla_enabled() or block_size == 1
            #                         else _Backend.TRITON_MLA)

            # if selected_backend == _Backend.TRITON_MLA:
            #     if block_size != 1:
zhuwenwen's avatar
zhuwenwen committed
238
239
240
241
242
243
244
245
            #         if use_v1:
            #             logger.info_once(
            #                 "Using Triton MLA backend on V1 engine.")
            #             return ("vllm.v1.attention.backends.mla."
            #                     "triton_mla.TritonMLABackend")
            #         else:
            #             logger.info("Using Triton MLA backend.")
            #             return "vllm.attention.backends.triton_mla.TritonMLABackend"  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
246
247
248
249
            #     else:
            #         raise ValueError(
            #             f" The selected backend, {selected_backend.name},"
            #             f"does not support block size {block_size}.")
zhuwenwen's avatar
zhuwenwen committed
250
251
            # elif selected_backend == _Backend.ROCM_AITER_MLA \
            #     or selected_backend == _Backend.ROCM_AITER_MLA_VLLM_V1:
zhuwenwen's avatar
zhuwenwen committed
252
            #     if block_size == 1:
zhuwenwen's avatar
zhuwenwen committed
253
254
255
256
257
258
            #         if use_v1:
            #             logger.info("Using AITER MLA backend on V1 engine.")
            #             return "vllm.v1.attention.backends.mla.rocm_aiter_mla.AiterMLABackend"  # noqa: E501
            #         else:
            #             logger.info("Using AITER MLA backend")
            #             return "vllm.attention.backends.rocm_aiter_mla.AiterMLABackend"  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
259
260
261
262
263
264
265
266
267
            #     else:
            #         raise ValueError(
            #             f" The selected backend, {selected_backend.name},"
            #             f"does not support block size {block_size}."
            #             "(currently only supports block size 1)")
            # else:
            #     raise ValueError(
            #         f" The selected backend, {selected_backend.name},"
            #         f"is not MLA type while requested for MLA backend.")
268
269
270
      
        if selected_backend is None or selected_backend == _Backend.FLASH_ATTN:
            selected_backend = _Backend.ROCM_FLASH
271

272
        if envs.VLLM_USE_V1:
zhuwenwen's avatar
zhuwenwen committed
273
274
275
276
277
278
            TRITON_ATTN_VLLM_V1 = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend"  # noqa: E501
            FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"  # noqa: E501
            # if selected_backend == _Backend.TRITON_ATTN_VLLM_V1:
            #     logger.info_once("Using Triton backend on V1 engine.")
            #     return TRITON_ATTN_VLLM_V1
            
279
            if envs.is_set("VLLM_USE_FLASH_ATTN_PA") and envs.VLLM_USE_FLASH_ATTN_PA and block_size == 64:
zhuwenwen's avatar
zhuwenwen committed
280
281
282
                logger.info_once("Using Flash Attention backend on V1 engine. (only supports block size 64)")
                return FLASH_ATTN_V1

283
            else:
zhuwenwen's avatar
zhuwenwen committed
284
285
286
                logger.info_once("Using Triton backend on V1 engine.")
                return TRITON_ATTN_VLLM_V1
            
287
288
289
290
291
292
293
294
        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)
        logger.info("Using ROCmFlashAttention backend.")
        return "vllm.attention.backends.rocm_flash_attn.ROCmFlashAttentionBackend"  # noqa: E501
295

296

297
298
299
300
301
302
303
    @classmethod
    def set_device(cls, device: torch.device) -> None:
        """
        Set the device for the current platform.
        """
        torch.cuda.set_device(device)

304
    @classmethod
305
    @lru_cache(maxsize=8)
306
307
308
    def get_device_capability(cls,
                              device_id: int = 0
                              ) -> Optional[DeviceCapability]:
309
310
        major, minor = torch.cuda.get_device_capability(device_id)
        return DeviceCapability(major=major, minor=minor)
311

312
    @classmethod
313
    @with_amdsmi_context
314
    def is_fully_connected(cls, physical_device_ids: list[int]) -> bool:
zhuwenwen's avatar
zhuwenwen committed
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
        """
        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
335

336
    @classmethod
337
    @with_amdsmi_context
338
    @lru_cache(maxsize=8)
339
    def get_device_name(cls, device_id: int = 0) -> str:
340
        physical_device_id = device_id_to_physical_device_id(device_id)
341
        handle = amdsmi_get_processor_handles()[physical_device_id]
342
343
        # return amdsmi_get_gpu_asic_info(handle)["market_name"]
        return torch.cuda.get_device_name(device_id)
344

zhuwenwen's avatar
zhuwenwen committed
345
346
347
348
    @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
349

350
351
    @classmethod
    def is_async_output_supported(cls, enforce_eager: Optional[bool]) -> bool:
352
        if enforce_eager and not envs.VLLM_USE_V1:
353
354
355
356
357
358
359
            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

360
    @classmethod
361
    def check_and_update_config(cls, vllm_config: "VllmConfig") -> None:
362
363
364
365
        cache_config = vllm_config.cache_config
        if cache_config and cache_config.block_size is None:
            cache_config.block_size = 16

366
367
368
369
        parallel_config = vllm_config.parallel_config
        scheduler_config = vllm_config.scheduler_config
        if parallel_config.worker_cls == "auto":
            if scheduler_config.is_multi_step:
370
371
372
                if envs.VLLM_USE_V1:
                    raise NotImplementedError(
                        "Multi-step scheduling is not supported (and not "
373
                        "needed) on vLLM V1. Please launch without "
374
375
376
377
                        "--num-scheduler-steps.")
                else:
                    parallel_config.worker_cls = \
                        "vllm.worker.multi_step_worker.MultiStepWorker"
378
            elif vllm_config.speculative_config:
379
                if not envs.VLLM_USE_V1:
380
                    raise NotImplementedError(
381
382
                        "Speculative decoding is not supported on vLLM V0.")
                parallel_config.worker_cls = "vllm.v1.worker.gpu_worker.Worker"
383
            else:
384
385
386
387
388
                if envs.VLLM_USE_V1:
                    parallel_config.worker_cls = \
                            "vllm.v1.worker.gpu_worker.Worker"
                else:
                    parallel_config.worker_cls = "vllm.worker.worker.Worker"
389

390
391
392
393
394
395
396
397
398
399
400
401
    @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)

402
403
404
405
406
407
    @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"
408
409
                " is not set, disabling VLLM_USE_TRITON_AWQ.")
            envs.VLLM_USE_TRITON_AWQ = False
410
411
412
413

    @classmethod
    def get_punica_wrapper(cls) -> str:
        return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU"
414
415
416
417
418
419

    @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
420
421
422
        # return torch.cuda.mem_get_info(device)[1] - torch.cuda.mem_get_info(
        #     device)[0]
        return torch.cuda.max_memory_allocated(device)
423
424
425
426

    @classmethod
    def get_device_communicator_cls(cls) -> str:
        return "vllm.distributed.device_communicators.cuda_communicator.CudaCommunicator"  # noqa
427

428
429
430
431
432
    @classmethod
    def supports_mx(cls) -> bool:
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
        return any(gfx in gcn_arch for gfx in ["gfx95"])

433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
    @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
449
450

    @classmethod
451
    def supports_v1(cls, model_config: "ModelConfig") -> bool:
452
453
        # V1 support on AMD gpus is experimental
        return True
454
455
456
457
458

    @classmethod
    def use_custom_allreduce(cls) -> bool:
        # We only enable custom allreduce for MI300 series
        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
459
        supported_archs = ['gfx94', 'gfx95']
460
        return any(gfx in gcn_arch for gfx in supported_archs)
461
462
463
464

    @classmethod
    def get_cu_count(cls, device_id: int = 0) -> int:
        return torch.cuda.get_device_properties(
465
            device_id).multi_processor_count
466
467
468
469

    @classmethod
    def is_navi(cls) -> bool:
        return 'gfx1' in torch.cuda.get_device_properties(0).gcnArchName
470
471
472
473

    @classmethod
    def get_piecewise_backend_cls(cls) -> str:
        return "vllm.compilation.cuda_piecewise_backend.CUDAPiecewiseBackend"  # noqa
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503

    @classmethod
    def stateless_init_device_torch_dist_pg(
        cls,
        backend: str,
        prefix_store: PrefixStore,
        group_rank: int,
        group_size: int,
        timeout: timedelta,
    ) -> ProcessGroup:
        assert is_nccl_available()
        pg: ProcessGroup = ProcessGroup(
            prefix_store,
            group_rank,
            group_size,
        )
        from torch.distributed.distributed_c10d import ProcessGroupNCCL

        backend_options = ProcessGroupNCCL.Options()
        backend_options._timeout = timeout

        backend_class = ProcessGroupNCCL(prefix_store, group_rank, group_size,
                                         backend_options)
        backend_type = ProcessGroup.BackendType.NCCL
        device = torch.device("cuda")
        pg._set_default_backend(backend_type)
        backend_class._set_sequence_number_for_group()

        pg._register_backend(device, backend_type, backend_class)
        return pg
504
505
506
507

    @classmethod
    def device_count(cls) -> int:
        return cuda_device_count_stateless()
508
509
510
511

    @classmethod
    def is_kv_cache_dtype_supported(cls, kv_cache_dtype: str) -> bool:
        return True