Skip to content
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

[LLM] [NPU] StaticLLMPipeline: Compiler DQ update #1515

Merged
merged 8 commits into from
Jan 14, 2025
Merged

[LLM] [NPU] StaticLLMPipeline: Compiler DQ update #1515

Changes from 4 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
ba70ef1
Update DQ query
smirnov-alexey Jan 9, 2025
f5dd5b1
Unconditionally utilize compiler DQ
smirnov-alexey Jan 13, 2025
cc44a0d
DQ only when in supported props
smirnov-alexey Jan 13, 2025
fa76cf7
Add prefix
smirnov-alexey Jan 13, 2025
868a7ac
Align DQ behaviour
smirnov-alexey Jan 13, 2025
3c72f4d
Address review comments
smirnov-alexey Jan 13, 2025
24b1a63
Merge branch 'master' into as/npuw_dq
smirnov-alexey Jan 13, 2025
102a1d9
Merge branch 'master' into as/npuw_dq
smirnov-alexey Jan 14, 2025
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
21 changes: 11 additions & 10 deletions src/cpp/src/llm_pipeline_static.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -475,11 +475,10 @@ std::optional<NPUDesc> extract_npu_descriptor(ov::Core& core) {
}
const auto arch = core.get_property("NPU", ov::device::architecture);
const auto max_tiles = core.get_property("NPU", ov::intel_npu::max_tiles);

bool compiler_dq = false;
const auto device_caps = core.get_property("NPU", ov::device::capabilities);
if (std::find(device_caps.begin(), device_caps.end(),
"COMPILER_DYNAMIC_QUANTIZATION") != device_caps.end()) {
const auto supported_properties = core.get_property("NPU", ov::supported_properties);
if (std::find(supported_properties.begin(), supported_properties.end(),
lmielick marked this conversation as resolved.
Show resolved Hide resolved
"NPU_COMPILER_DYNAMIC_QUANTIZATION") != supported_properties.end()) {
compiler_dq = true;
}
return std::make_optional(NPUDesc{arch, max_tiles, compiler_dq});
Expand Down Expand Up @@ -516,6 +515,10 @@ ov::AnyMap get_default_prefill_config(const std::shared_ptr<ov::Model>& model,
auto config = get_default_common_config(model);
if (is_cw_compressed(model)) {
config.emplace("NPUW_DQ", "YES");
if (npudesc.has_value() && npudesc->compiler_dq) {
config.emplace("NPUW_DQ_FULL", "NO");
config.emplace("NPU_COMPILER_DYNAMIC_QUANTIZATION", true);
}
Copy link
Contributor

@dmatveev dmatveev Jan 13, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why do you enable it for CW compressed models only? The whole point of this feature was to make it for group-quantized prefill.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

} else {
config.emplace("NPUW_PMM", "NO");
}
Expand All @@ -524,9 +527,6 @@ ov::AnyMap get_default_prefill_config(const std::shared_ptr<ov::Model>& model,
npudesc->max_tiles != -1) {
config.emplace("NPU_DPU_GROUPS", npudesc->max_tiles);
}
if (npudesc.has_value() && npudesc->compiler_dq) {
config.emplace("NPUW_DQ_FULL", "NO");
}
return config;
}

Expand All @@ -539,15 +539,16 @@ ov::AnyMap get_default_generate_config(const std::shared_ptr<ov::Model>& model,
}
// NB: Unconditionally set for generation model
config.emplace("NPUW_DQ", "YES");
if (npudesc.has_value() && npudesc->compiler_dq) {
config.emplace("NPUW_DQ_FULL", "NO");
config.emplace("NPU_COMPILER_DYNAMIC_QUANTIZATION", true);
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You certainly don't need to make it twice, just do this in the shared section of the config (so it goes to both prefill and kvcache).

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also, when you switch to compiler DQ, you'll have to disable DCOFF since otherwise DCOFF will be applied to this model and it will run in FP16.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

if (npudesc.has_value() && npudesc->arch == "4000") {
config.emplace("NPU_DPU_GROUPS", 4);
}
if (hint == GenerateHint::FAST_COMPILE) {
config.emplace("NPUW_UNFOLD_IREQS", "YES");
}
if (npudesc.has_value() && npudesc->compiler_dq) {
config.emplace("NPUW_DQ_FULL", "NO");
}
return config;
}

Expand Down
Loading