Skip to content

Fix for view+inplace error from Falcon #127

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
vivekgoe merged 11 commits into from
Mar 21, 2024
Merged

Fix for view+inplace error from Falcon #127

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
c72250a
enable Falcon FP8 inference
Mar 6, 2024
63fd6b2
added example command in readme, code cleanup
Mar 7, 2024
af025a7
resolve issues in finetuning
Mar 8, 2024
7b9852a
enable non reuse cache flow for fp8
Mar 11, 2024
8a18736
revert non reuse_cache flow for training due to perf drop
Mar 13, 2024
3feaed0
add falcon180B FP8 test
Mar 13, 2024
f849be5
Merge branch 'habana-main' into schoi/falcon_180b_quant_PR
Mar 14, 2024
1b90b33
fix error
Mar 14, 2024
5a2128b
fix run_lm_eval.py to save --reuse_cache
schoi-habana Mar 16, 2024
8101af0
Merge branch 'habana-main' into schoi/falcon_180b_quant_PR
schoi-habana Mar 20, 2024
1afafdf
fix Falcon view+inplace error
schoi-habana Mar 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion examples/text-generation/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -240,7 +240,7 @@ def setup_distributed_model(args, model_dtype, model_kwargs, logger):

model = deepspeed.init_inference(model, **ds_inference_kwargs)
model = model.module
if model.config.model_type == "llama" or "falcon":
if model.config.model_type in ["llama", "falcon"]:
patch_scoped_linear_all_reduce(model)

if args.quant_config:
Expand Down
39 changes: 22 additions & 17 deletions optimum/habana/transformers/models/falcon/modeling_falcon.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,9 +65,10 @@ def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training:
out = F.dropout(x, p=prob, training=training)
if training:
out = residual + out
return out
else:
out.add_(residual)
return out
residual.add_(out)
return residual


def apply_customized_rope(q, k, cos, sin, position_ids):
Expand Down Expand Up @@ -536,21 +537,25 @@ def forward(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
residual = hidden_states
hidden_states, present, attn_scores, attention_layernorm_out, mlp_layernorm_out = (
self.pre_attn( # layernorm + attention before AllReduce
hidden_states,
layer_past=layer_past,
attention_mask=attention_mask,
position_ids=position_ids,
alibi=alibi,
head_mask=head_mask,
use_cache=use_cache,
output_attentions=output_attentions,
token_idx=token_idx,
reuse_cache=reuse_cache,
cache_idx=cache_idx,
**kwargs,
)
(
hidden_states,
present,
attn_scores,
attention_layernorm_out,
mlp_layernorm_out,
) = self.pre_attn( # layernorm + attention before AllReduce
hidden_states,
layer_past=layer_past,
attention_mask=attention_mask,
position_ids=position_ids,
alibi=alibi,
head_mask=head_mask,
use_cache=use_cache,
output_attentions=output_attentions,
token_idx=token_idx,
reuse_cache=reuse_cache,
cache_idx=cache_idx,
**kwargs,
)

self.self_attention.attention_all_reduce(hidden_states)
Expand Down