Revert "Enable Nemotron-H FP8 on HPU: quant-aware Mamba in_proj + non-gated FP8 MoE (#1666)" - #1785
Open
rsmyrek wants to merge 1 commit into
Conversation
…-gated FP8 MoE (vllm-project#1666)" This reverts commit 058e248. Signed-off-by: Radoslaw Smyrek <radoslawx.smyrek@intel.com>
Contributor
There was a problem hiding this comment.
Warning
Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.
Pull request overview
Reverts the prior Nemotron-H FP8 enablement on HPU by removing Nemotron-H-specific logic across Mamba KV cache grouping, Mamba mixer behavior, and MoE activation/platform patches.
Changes:
- Removes Nemotron-H from model feature flags and Mamba-like architecture handling.
- Drops Nemotron-H / FP8-quant-aware paths in Mamba mixer and mixed-dtype Mamba KV-cache allocation logic.
- Removes non-gated MoE activation aliasing and the HPU-side workaround patch for vLLM core’s non-gated MoE guard.
Reviewed changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| vllm_gaudi/v1/worker/hpu_model_runner.py | Reverts Nemotron-H handling in Mamba-related KV cache group assignment and mixed-dtype KV-cache allocation. |
| vllm_gaudi/ops/hpu_mamba_mixer2.py | Removes quant-aware in_proj path and flatten_input early-return behavior. |
| vllm_gaudi/ops/hpu_fused_moe.py | Removes activation aliasing for Nemotron-H and deletes the patched non-gated MoE config workaround. |
| vllm_gaudi/ops/hpu_compressed_tensors.py | Removes non-gated MoE plumbing and simplifies activation normalization usage. |
| vllm_gaudi/extension/ops.py | Removes is_gated state and the conditional is_gated kwarg forwarding to the HPU MoE op. |
| vllm_gaudi/extension/features.py | Removes nemotron_h from flatten_input feature enablement. |
Suppressed comments (1)
vllm_gaudi/ops/hpu_fused_moe.py:1
- Removing the
"relu2_no_mul" -> "relu2"alias can turn a previously accepted activation name into an unknown string passed to the HPU custom op. Even if Nemotron-H is being reverted, keeping this alias is low-risk and improves robustness/backward-compatibility for configs that might still surface"relu2_no_mul"(e.g., cached configs, external integrations). Consider restoring the alias mapping unless you can guarantee it will never be used."
from collections.abc import Callable
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+305
to
+307
| states_proj = F.linear(hidden_states, self._states_weight, self._states_bias) | ||
|
|
||
| gate = F.linear(hidden_states, self._gate_weight, self._gate_bias) |
Comment on lines
+820
to
+821
| parts = layer_name.split('.') | ||
| layer_idx = int(parts[-2]) # "model.layers.5.mixer" -> 5 |
Comment on lines
+825
to
+826
| layer = model.model.layers[layer_idx] | ||
| layer.mamba.cache_group_idx = group_idx |
| @@ -343,14 +330,6 @@ def forward( | |||
| # 5. Final linear projection | |||
| output, _ = self.out_proj(hidden_states_varlen) | |||
|
|
|||
| @@ -6667,15 +6656,7 @@ def _needs_raw_buffer(kv_cache_tensor) -> bool: | |||
| if isinstance(spec, MambaSpec) and \ | |||
| spec.mamba_type in _GDN_MAMBA_TYPES: | |||
| continue | |||
Contributor
✅ CI PassedAll checks passed successfully against the following vllm commit: |
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
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.
Enablement moved to next release.
This reverts commit 058e248.