File size: 6,573 Bytes
d287f83 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import pytest
import torch
from loguru import logger
from transformers.models.mixtral.modeling_mixtral import MixtralDecoderLayer, MixtralRotaryEmbedding
import ttnn
from models.common.utility_functions import comp_allclose, comp_pcc
from models.tt_transformers.tt.ccl import TT_CCL
from models.tt_transformers.tt.common import Mode
from models.tt_transformers.tt.decoder import TransformerBlock as TtTransformerBlock
from models.tt_transformers.tt.model_config import ModelArgs
from models.tt_transformers.tt.rope import RotarySetup
from .utils import fuse_mixtral_experts, load_hf_mixtral_config
# pytest models/tt_transformers/tests/mixtral/test_mixtral_decoder.py
def convert2ref(state_dict):
"""Convert state dict keys for compatibility with reference model naming."""
replacements = [
("attention.wq.weight", "self_attn.q_proj.weight"),
("attention.wk.weight", "self_attn.k_proj.weight"),
("attention.wv.weight", "self_attn.v_proj.weight"),
("attention.wo.weight", "self_attn.o_proj.weight"),
("attention_norm.weight", "input_layernorm.weight"),
("ffn_norm.weight", "post_attention_layernorm.weight"),
# transformers 5.x renamed MixtralDecoderLayer.block_sparse_moe -> mlp
("block_sparse_moe.gate.weight", "mlp.gate.weight"),
]
out = {}
for k, v in state_dict.items():
new_k = k
for old, new in replacements:
if k.startswith(old):
new_k = k.replace(old, new)
break
out[new_k] = v
return out
@pytest.mark.parametrize("layer_idx", [0])
@pytest.mark.parametrize(
"batch",
(32,),
)
@pytest.mark.parametrize("device_params", [{"fabric_config": True}], indirect=True)
@pytest.mark.parametrize("mesh_device", [(1, 8)], indirect=True)
def test_mixtral_decoder_inference(mesh_device, reset_seeds, batch, device_params, layer_idx):
"""
b: batch
s: sequence length
h: hidden size
"""
pcc = 0.98
dtype = ttnn.bfloat8_b
if batch == 32:
generation_start_pos = 15000
max_seq_len = 16384
elif batch in [4, 8, 16]:
generation_start_pos = 30000
max_seq_len = 32768
else:
raise ValueError(f"Batch size {batch} not supported")
hf_config = load_hf_mixtral_config()
model_args = ModelArgs(mesh_device, max_seq_len=max_seq_len, max_batch_size=batch)
model_args.use_qk_fused = False
state_dict = model_args.load_state_dict()
partial_state_dict = {k[9:]: v for k, v in state_dict.items() if (k.startswith(f"layers.{layer_idx}."))}
reference_model = MixtralDecoderLayer(hf_config, layer_idx)
# transformers 5.x: experts are fused into mlp.experts.{gate_up_proj,down_proj}
reference_model.load_state_dict(
fuse_mixtral_experts(
convert2ref(partial_state_dict), hf_config.num_local_experts, "block_sparse_moe.experts", "mlp.experts"
)
)
reference_rotary_emb = MixtralRotaryEmbedding(config=hf_config)
# Initialize TT model
rope_setup = RotarySetup(
mesh_device,
model_args.max_batch_size,
model_args.head_dim,
model_args.max_seq_len,
model_args.rope_theta,
model_args.rope_scaling,
)
tt_ccl = TT_CCL(mesh_device)
tt_model = TtTransformerBlock(
mesh_device=mesh_device,
state_dict=state_dict,
args=model_args,
layer_num=layer_idx,
dtype=dtype,
weight_cache_path=model_args.weight_cache_path(dtype),
transformation_mats=rope_setup.get_both_trans_mats(),
tt_ccl=tt_ccl,
)
generation_length = 10
all_tests_pass = True
seqlen = 1
for i in range(generation_length):
logger.info(f"[Decoder] Generating token {i}")
pt_decode_input_bsh = (torch.rand(batch, seqlen, model_args.dim) * 2) - 1
start_pos = generation_start_pos + i
start_pos_ids = torch.tensor([start_pos for _ in range(batch)])
current_pos_tensor = ttnn.from_torch(
start_pos_ids,
device=mesh_device,
dtype=ttnn.int32,
mesh_mapper=ttnn.ShardTensor2dMesh(
mesh_device,
dims=(None, None),
mesh_shape=model_args.cluster_shape,
),
)
tt_decode_input = pt_decode_input_bsh.clone()
decode_input = model_args.prepare_residual_tensor_decode(
tt_decode_input,
model_args.get_residual_mem_config(Mode.DECODE, None),
)
rot_mats = rope_setup.get_rot_mats(start_pos_ids)
# Run TT model
tt_out_b1sh = tt_model(
decode_input,
current_pos_tensor,
rot_mats_global=rot_mats,
mode=Mode.DECODE,
)
tt_out = (
ttnn.to_torch(
tt_out_b1sh,
mesh_composer=ttnn.ConcatMesh2dToTensor(mesh_device, dims=(1, 3), mesh_shape=model_args.cluster_shape),
)[0]
.squeeze(0)
.view(batch, 1, -1)
)
# Reference model
positions = torch.LongTensor([start_pos])
position_ids = positions.unsqueeze(0)
position_embeddings = reference_rotary_emb(pt_decode_input_bsh, position_ids)
# transformers 5.x MixtralDecoderLayer.forward returns a single tensor (older versions
# returned a tuple). `ref_output_bsh, *_ =` would unpack the tensor along the batch dim,
# so assign directly and take element 0 only when a tuple/list is returned.
ref_output_bsh = reference_model(
hidden_states=pt_decode_input_bsh,
position_ids=position_ids,
position_embeddings=position_embeddings,
)
if isinstance(ref_output_bsh, (tuple, list)):
ref_output_bsh = ref_output_bsh[0]
passing, pcc_message = comp_pcc(ref_output_bsh, tt_out, pcc)
logger.info(comp_allclose(ref_output_bsh, tt_out))
logger.info(pcc_message)
if passing:
logger.info("Mixtral Decoder Block Passed!")
else:
logger.warning("Mixtral Decoder Block Failed!")
all_tests_pass = False
if all_tests_pass:
logger.info(f"All {generation_length} Mixtral decode iterations Passed!")
else:
logger.warning("One or more iterations of Mixtral decode Failed!")
assert all_tests_pass, f"PCC value is lower than {pcc} for some of the outputs. Check Warnings!"
|