mosaicml

/

mpt-7b-instruct

like 460

I'm loading in the triton implementation of the model using a custom device map and trying to generate an output as follows (to be clear, I have no issues with the torch implementation):

torch_dtype = torch.bfloat16

config = AutoConfig.from_pretrained(
  'mosaicml/mpt-7b',
    trust_remote_code=True
)
config.attn_config['attn_impl'] = 'triton'
config.update({"max_seq_len": max_len})
config.update({"torch_dtype": torch_dtype})

with open('MPT_device_map.pkl', 'rb') as f:
    dm = pickle.load(f)

model = AutoModelForCausalLM.from_pretrained(
    'mosaicml/mpt-7b-instruct',
    torch_dtype=torch_dtype,
    trust_remote_code=True,
    device_map=dm,
    config=config,
    local_files_only=True
)

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b", padding_side="left")

tokenizer.pad_token = tokenizer.eos_token

inputs = tokenizer(text, return_tensors="pt", padding=True).input_ids.to(device)

streamer = TextStreamer(tokenizer)

with torch.inference_mode():
    generate_ids = model.generate(inputs, **params, streamer=streamer)
    generate_ids = generate_ids[:,inputs[0].shape[-1]:]
    output = tokenizer.batch_decode(generate_ids, skip_special_tokens=True)[0]

And I'm getting the following error:
```
/usr/bin/ld: skipping incompatible /usr/lib/libcuda.so when searching for -lcuda

KeyError Traceback (most recent call last)
File :21, in _fwd_kernel(Q, K, V, Bias, Out, Lse, TMP, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_ob, stride_oh, stride_om, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE, IS_CAUSAL, BLOCK_HEADDIM, EVEN_M, EVEN_N, EVEN_HEADDIM, BLOCK_M, BLOCK_N, grid, num_warps, num_stages, extern_libs, stream, warmup)

KeyError: ('2-.-0-.-0-d82511111ad128294e9d31a6ac684238-2b0c5161c53c71b37ae20a9996ee4bb8-c1f92808b4e4644c1732e8338187ac87-d962222789c30252d492a16cca3bf467-12f7ac1ca211e037f62a7c0c323d9990-5c5e32ff210f3b7f56c98ca29917c25e-06f0df2d61979d629033f4a22eff5198-0dd03b0bd512a184b3512b278d9dfa59-d35ab04ae841e2714a253c523530b071', (torch.bfloat16, torch.bfloat16, torch.bfloat16, torch.bfloat16, torch.bfloat16, torch.float32, torch.float32, 'fp32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32', 'i32'), ('vector', True, 128, False, False, True, 128, 128), (True, True, True, True, True, True, True, (False,), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False), (True, False)))

During handling of the above exception, another exception occurred:

RuntimeError Traceback (most recent call last)
Cell In[41], line 8
4 with torch.inference_mode():
5 # res = nlp(text, max_new_tokens=mnt, min_new_tokens=1, return_full_text=False)
6 # inputs = input_map(inputs)
7 st = time.time()
----> 8 generate_ids = model.generate(inputs, **params, streamer=streamer)
9 # generate_ids = model.module.generate(**inputs, **params, streamer=streamer)
10 tt = time.time() - st

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/torch/autograd/grad_mode.py:27, in _DecoratorContextManager.call..decorate_context(*args, **kwargs)
24 @functools.wraps(func)
25 def decorate_context(*args, **kwargs):
26 with self.clone():
---> 27 return func(*args, **kwargs)

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/transformers/generation/utils.py:1565, in GenerationMixin.generate(self, inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, assistant_model, streamer, **kwargs)
1557 input_ids, model_kwargs = self._expand_inputs_for_generation(
1558 input_ids=input_ids,
1559 expand_size=generation_config.num_return_sequences,
1560 is_encoder_decoder=self.config.is_encoder_decoder,
1561 **model_kwargs,
1562 )
1564 # 13. run sample
-> 1565 return self.sample(
1566 input_ids,
1567 logits_processor=logits_processor,
1568 logits_warper=logits_warper,
1569 stopping_criteria=stopping_criteria,
1570 pad_token_id=generation_config.pad_token_id,
1571 eos_token_id=generation_config.eos_token_id,
1572 output_scores=generation_config.output_scores,
1573 return_dict_in_generate=generation_config.return_dict_in_generate,
1574 synced_gpus=synced_gpus,
1575 streamer=streamer,
1576 **model_kwargs,
1577 )
1579 elif is_beam_gen_mode:
1580 if generation_config.num_return_sequences > generation_config.num_beams:

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/transformers/generation/utils.py:2612, in GenerationMixin.sample(self, input_ids, logits_processor, stopping_criteria, logits_warper, max_length, pad_token_id, eos_token_id, output_attentions, output_hidden_states, output_scores, return_dict_in_generate, synced_gpus, streamer, **model_kwargs)
2609 model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
2611 # forward pass to get next token
-> 2612 outputs = self(
2613 **model_inputs,
2614 return_dict=True,
2615 output_attentions=output_attentions,
2616 output_hidden_states=output_hidden_states,
2617 )
2619 if synced_gpus and this_peer_finished:
2620 continue # don't waste resources running the code we don't need

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/torch/nn/modules/module.py:1194, in Module._call_impl(self, *input, **kwargs)
1190 # If we don't have any hooks, we want to skip the rest of the logic in
1191 # this function, and just call forward.
1192 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1193 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1194 return forward_call(*input, **kwargs)
1195 # Do not call functions when jit is used
1196 full_backward_hooks, non_full_backward_hooks = [], []

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/accelerate/hooks.py:165, in add_hook_to_module..new_forward(*args, **kwargs)
163 output = old_forward(*args, **kwargs)
164 else:
--> 165 output = old_forward(*args, **kwargs)
166 return module._hf_hook.post_forward(module, output)

File ~/.cache/huggingface/modules/transformers_modules/mosaicml/mpt-7b/d8304854d4877849c3c0a78f3469512a84419e84/modeling_mpt.py:237, in MPTForCausalLM.forward(self, input_ids, past_key_values, attention_mask, prefix_mask, sequence_id, labels, return_dict, output_attentions, output_hidden_states, use_cache)
235 return_dict = return_dict if return_dict is not None else self.config.return_dict
236 use_cache = use_cache if use_cache is not None else self.config.use_cache
--> 237 outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache)
238 logits = F.linear(outputs.last_hidden_state, self.transformer.wte.weight)
239 if self.logit_scale is not None:

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/torch/nn/modules/module.py:1194, in Module._call_impl(self, *input, **kwargs)
1190 # If we don't have any hooks, we want to skip the rest of the logic in
1191 # this function, and just call forward.
1192 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1193 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1194 return forward_call(*input, **kwargs)
1195 # Do not call functions when jit is used
1196 full_backward_hooks, non_full_backward_hooks = [], []

File ~/.cache/huggingface/modules/transformers_modules/mosaicml/mpt-7b/d8304854d4877849c3c0a78f3469512a84419e84/modeling_mpt.py:183, in MPTModel.forward(self, input_ids, past_key_values, attention_mask, prefix_mask, sequence_id, return_dict, output_attentions, output_hidden_states, use_cache)
181 all_hidden_states = all_hidden_states + (x,)
182 past_key_value = past_key_values[b_idx] if past_key_values is not None else None
--> 183 (x, past_key_value) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=self.is_causal)
184 if past_key_values is not None:
185 past_key_values[b_idx] = past_key_value

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/torch/nn/modules/module.py:1194, in Module._call_impl(self, *input, **kwargs)
1190 # If we don't have any hooks, we want to skip the rest of the logic in
1191 # this function, and just call forward.
1192 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1193 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1194 return forward_call(*input, **kwargs)
1195 # Do not call functions when jit is used
1196 full_backward_hooks, non_full_backward_hooks = [], []

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/accelerate/hooks.py:165, in add_hook_to_module..new_forward(*args, **kwargs)
163 output = old_forward(*args, **kwargs)
164 else:
--> 165 output = old_forward(*args, **kwargs)
166 return module._hf_hook.post_forward(module, output)

File ~/.cache/huggingface/modules/transformers_modules/mosaicml/mpt-7b/d8304854d4877849c3c0a78f3469512a84419e84/blocks.py:36, in MPTBlock.forward(self, x, past_key_value, attn_bias, attention_mask, is_causal)
34 def forward(self, x: torch.Tensor, past_key_value: Optional[Tuple[torch.Tensor]]=None, attn_bias: Optional[torch.Tensor]=None, attention_mask: Optional[torch.ByteTensor]=None, is_causal: bool=True) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
35 a = self.norm_1(x)
---> 36 (b, _, past_key_value) = self.attn(a, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=is_causal)
37 x = x + self.resid_attn_dropout(b)
38 m = self.norm_2(x)

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/torch/nn/modules/module.py:1194, in Module._call_impl(self, *input, **kwargs)
1190 # If we don't have any hooks, we want to skip the rest of the logic in
1191 # this function, and just call forward.
1192 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1193 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1194 return forward_call(*input, **kwargs)
1195 # Do not call functions when jit is used
1196 full_backward_hooks, non_full_backward_hooks = [], []

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/accelerate/hooks.py:165, in add_hook_to_module..new_forward(*args, **kwargs)
163 output = old_forward(*args, **kwargs)
164 else:
--> 165 output = old_forward(*args, **kwargs)
166 return module._hf_hook.post_forward(module, output)

File ~/.cache/huggingface/modules/transformers_modules/mosaicml/mpt-7b/d8304854d4877849c3c0a78f3469512a84419e84/attention.py:171, in MultiheadAttention.forward(self, x, past_key_value, attn_bias, attention_mask, is_causal, needs_weights)
169 if attn_bias is not None:
170 attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
--> 171 (context, attn_weights) = self.attn_fn(query, key, value, self.n_heads, softmax_scale=self.softmax_scale, attn_bias=attn_bias, key_padding_mask=key_padding_mask, is_causal=is_causal, dropout_p=self.attn_dropout_p, training=self.training, needs_weights=needs_weights)
172 return (self.out_proj(context), attn_weights, past_key_value)

File ~/.cache/huggingface/modules/transformers_modules/mosaicml/mpt-7b/d8304854d4877849c3c0a78f3469512a84419e84/attention.py:111, in triton_flash_attn_fn(query, key, value, n_heads, softmax_scale, attn_bias, key_padding_mask, is_causal, dropout_p, training, needs_weights, multiquery)
109 value = value.expand(*value.shape[:2], n_heads, value.size(-1))
110 reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
--> 111 attn_output = flash_attn_triton.flash_attn_func(query, key, value, attn_bias, reset_is_causal, softmax_scale)
112 output = attn_output.view(*attn_output.shape[:2], -1)
113 return (output, None)

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/flash_attn/flash_attn_triton.py:810, in FlashAttnFunc.forward(ctx, q, k, v, bias, causal, softmax_scale)
808 # Make sure that the last dimension is contiguous
809 q, k, v = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, k, v]]
--> 810 o, lse, ctx.softmax_scale = _flash_attn_forward(
811 q, k, v, bias=bias, causal=causal, softmax_scale=softmax_scale
812 )
813 ctx.save_for_backward(q, k, v, o, lse, bias)
814 ctx.causal = causal

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/flash_attn/flash_attn_triton.py:623, in _flash_attn_forward(q, k, v, bias, causal, softmax_scale)
621 num_warps = 4 if d <= 64 else 8
622 grid = lambda META: (triton.cdiv(seqlen_q, META["BLOCK_M"]), batch * nheads)
--> 623 _fwd_kernel[grid](
624 q, k, v, bias, o,
625 lse, tmp,
626 softmax_scale,
627 q.stride(0), q.stride(2), q.stride(1),
628 k.stride(0), k.stride(2), k.stride(1),
629 v.stride(0), v.stride(2), v.stride(1),
630 *bias_strides,
631 o.stride(0), o.stride(2), o.stride(1),
632 nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d,
633 seqlen_q // 32, seqlen_k // 32, # key for triton cache (limit number of compilations)
634 # Can't use kwargs here because triton autotune expects key to be args, not kwargs
635 # IS_CAUSAL=causal, BLOCK_HEADDIM=d,
636 bias_type, causal, BLOCK_HEADDIM,
637 BLOCK_M=BLOCK, BLOCK_N=BLOCK,
638 num_warps=num_warps,
639 num_stages=1,
640 )
641 return o, lse, softmax_scale

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/triton/runtime/jit.py:106, in KernelInterface.getitem..launcher(*args, **kwargs)
105 def launcher(*args, **kwargs):
--> 106 return self.run(*args, grid=grid, **kwargs)

File ~/anaconda3/envs/llm/lib/python3.10/site-packages/triton/runtime/autotuner.py:200, in Heuristics.run(self, *args, **kwargs)
198 for v, heur in self.values.items():
199 kwargs[v] = heur({**dict(zip(self.arg_names, args)), **kwargs})
--> 200 return self.fn.run(*args, **kwargs)

File :43, in _fwd_kernel(Q, K, V, Bias, Out, Lse, TMP, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_ob, stride_oh, stride_om, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE, IS_CAUSAL, BLOCK_HEADDIM, EVEN_M, EVEN_N, EVEN_HEADDIM, BLOCK_M, BLOCK_N, grid, num_warps, num_stages, extern_libs, stream, warmup)

RuntimeError: Triton Error [CUDA]: invalid argument

Any ideas what might be causing this? I'm working with:
Python: 3.10
CUDA: 11.7
triton: 2.0.0.dev20221202
flash-attn: 1.0.3.post0
transformers: 4.29.2
torch: 1.13.1+cu117