Qualcomm AI Engine Direct - Model sharding for LLM #4923
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/4923
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 777fa22 with merge base 3fb03dc ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
Hi @cccclai, |
cccclai
reviewed
Aug 27, 2024
cccclai
reviewed
Aug 27, 2024
cccclai
reviewed
Aug 27, 2024
cccclai
approved these changes
Aug 27, 2024
Thanks for your prompt review. It should be work for llama model now.
class ReplaceIndexPutInput(ExportPass):
"""
Index put input workaround for quantized module
"""
dq_q_map = {
# per tensor
exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default: exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
exir_ops.edge.quantized_decomposed.dequantize_per_tensor.tensor: exir_ops.edge.quantized_decomposed.quantize_per_tensor.tensor,
# per channel
exir_ops.edge.quantized_decomposed.dequantize_per_channel.default: exir_ops.edge.quantized_decomposed.quantize_per_channel.default,
}
def __init__(self, edge_program: torch.export.ExportedProgram):
super(ReplaceIndexPutInput, self).__init__()
self.edge_program = edge_program
def call(self, graph_module: torch.fx.GraphModule):
graph = graph_module.graph
for node in graph.nodes:
if node.target == exir_ops.edge.aten.index_put.default:
if (copy_node := list(node.users)[0]) and copy_node.target == exir_ops.edge.aten.copy.default:
m_buffer_node = copy_node.args[0]
bad_frozen_node = node.args[0]
if QCOM_QUANT_ATTRS in bad_frozen_node.meta:
m_buffer_node.meta[QCOM_QUANT_ATTRS] = bad_frozen_node.meta[QCOM_QUANT_ATTRS]
m_buffer_node.meta[QCOM_QUANT_ATTRS][QCOM_ENCODING] = self.dq_q_map[m_buffer_node.meta[QCOM_QUANT_ATTRS][QCOM_ENCODING]]
with graph.inserting_after(bad_frozen_node):
node.replace_input_with(bad_frozen_node, m_buffer_node)
else:
continue
graph.eliminate_dead_code()
graph_module.recompile()
return PassResult(graph_module, True)
|
shewu-quic
force-pushed
the
dev1/hutton/model_sharding_with_custom_op
branch
2 times, most recently
from
19adc79 to
20d1e12
Compare
|
Hello team, I was trying to quantize the Llama 3.1 8B model against the QNN backend. After pulling in this PR, I had this issue below: |
Oops, Sorry about that "get_n_layers" seems to be removed from meta data. Let me fix it. |
For LLM, model size is too large to fit in device memory for inference. Therefore, we need to divide the model into a few parts in order to avoid inference time out-of-memory errors. Summary: - Use custom fallback op to split graph - Add splill fill feature - Add model sharding argument for qnn
shewu-quic
force-pushed
the
dev1/hutton/model_sharding_with_custom_op
branch
from
20d1e12 to
777fa22
Compare
I can help with that. Sorry I don’t have access to my regular machines recently and had a difficult time to repro the qnn flow on Ubuntu Linux (wsl) system. But this change should be simple enough and doesn’t require setting up qnn dependencies |
|
Just want to confirm on this
Any specific reason it's needed? Is it because the graph is better? |
The past kv cache is eliminated (frozen) with capture_pre_autograd_graph after convet_pt2e(fold_quantize=True). |
Did you expect changes like this #4942? Also wanted to check, if we replace |
Yes, it is similar to our change, but we don't set "strict". Does it affect anything? |
|
Hi @cccclai, @shewu-quic, I am trying to load the llama 3.1 8B model sharded in 4 to the QNN example. But the loading failed with memory buffer issues. The same workflow works for llama 2 7B model without using any sharding. I wonder if it is related to this diff and there are things needed to update for the model loading on-device? This is on OnePlus 12 with 16GB RAM, so memory size shouldn't be the issue. Here is the full log |
Hi @WuhanMonkey, if you saw exact pdId 2 Then it's related to the system. It's hard to do anything on the application side. I can only suggest
|
Thank you. Beside OnePlus phone, can it work on S24+ without any issue? |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For LLM, model size is too large to fit in device memory for inference. Therefore, we need to divide the model into a few parts in order to avoid inference time out-of-memory errors.
Summary: