Skip to content

Add settings to gitignore - #16

Merged
justinchuby merged 1 commit into
mainfrom
justinchu/ignore
Mar 20, 2026
Merged

Add settings to gitignore#16
justinchuby merged 1 commit into
mainfrom
justinchu/ignore

Conversation

@justinchuby

@justinchuby justinchuby commented Mar 20, 2026

Copy link
Copy Markdown
Member

No description provided.

@justinchuby
justinchuby requested a review from a team March 20, 2026 18:40
@justinchuby
justinchuby merged commit 4cfe763 into main Mar 20, 2026
5 of 14 checks passed
@justinchuby
justinchuby deleted the justinchu/ignore branch March 20, 2026 18:40
@github-actions

github-actions Bot commented Mar 20, 2026

Copy link
Copy Markdown

Performance Comparison

Model Metric Baseline Current Delta
falcon model_size_bytes 364 KB 364 KB +0.0% ⚪
falcon num_nodes 66 66 +0.0% ⚪
gemma2 model_size_bytes 428 KB 428 KB +0.0% ⚪
gemma2 num_nodes 109 107 -1.8% ⚪
gpt2 model_size_bytes 388 KB 388 KB +0.0% ⚪
gpt2 num_nodes 60 60 +0.0% ⚪
llama model_size_bytes 425 KB 425 KB +0.0% ⚪
llama num_nodes 68 68 +0.0% ⚪
llama (static-cache-text-generation) model_size_bytes 425 KB 425 KB +0.0% ⚪
llama (static-cache-text-generation) num_nodes 58 58 +0.0% ⚪
phi3 model_size_bytes 421 KB 421 KB +0.0% ⚪
phi3 num_nodes 68 68 +0.0% ⚪
phi3 (static-cache-text-generation) model_size_bytes 421 KB 421 KB +0.0% ⚪
phi3 (static-cache-text-generation) num_nodes 58 58 +0.0% ⚪
qwen2 model_size_bytes 425 KB 425 KB +0.0% ⚪
qwen2 num_nodes 68 68 +0.0% ⚪
qwen2 (static-cache-text-generation) model_size_bytes 425 KB 425 KB +0.0% ⚪
qwen2 (static-cache-text-generation) num_nodes 58 58 +0.0% ⚪

No performance regressions.

gramalingam added a commit that referenced this pull request Mar 23, 2026
@gramalingam
Document dtype-agnostic CastLike pattern in skills
e0ada90 
Add troubleshooting entry #16 to adding-a-new-model skill covering the
hardcoded Cast(to=float32) pitfall that breaks bfloat16/float16 models.

Add design principle #7 (dtype-agnostic components) and a CastLike op
pattern section to the reusable-components skill, explaining when to use
CastLike vs Cast.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 23, 2026
@gramalingam
Upcast SSM recurrence and gated norms to fp32 for precision parity
dbd4564 
Match HuggingFace's precision behaviour for numerically sensitive ops:
- SelectiveScan: upcast dt (softplus), A_log (exp), B, C, x, and state
  update to float32; cast output and new_ssm_state back to input dtype.
- Mamba2Scan: same fp32 upcast for the multi-head SSM recurrence.
- GatedRMSNorm: upcast both hidden_states and gate to float32 for
  SiLU gating and variance computation, matching Zamba2RMSNormGated.
- PostGatedRMSNorm: use stash_type=1 for RMSNorm, fp32 for SiLU gate.
- GatedDeltaNet: upcast softplus/exp decay computation to float32.
- _RMSNorm (SSM internal): upcast variance computation to float32.

Update skills documentation:
- reusable-components: replace CastLike-only guidance with complete
  fp32 upcast pattern (upcast → compute → cast back), table of which
  ops need fp32, and when upcast is NOT needed.
- adding-a-new-model: rewrite troubleshooting #16 to cover both type
  mismatch and numerical drift, with the correct two-part pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 23, 2026
@gramalingam
Document dtype-agnostic CastLike pattern in skills
c9e9e0c 
Add troubleshooting entry #16 to adding-a-new-model skill covering the
hardcoded Cast(to=float32) pitfall that breaks bfloat16/float16 models.

Add design principle #7 (dtype-agnostic components) and a CastLike op
pattern section to the reusable-components skill, explaining when to use
CastLike vs Cast.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 23, 2026
@gramalingam
Upcast SSM recurrence and gated norms to fp32 for precision parity
f181ef7 
Match HuggingFace's precision behaviour for numerically sensitive ops:
- SelectiveScan: upcast dt (softplus), A_log (exp), B, C, x, and state
  update to float32; cast output and new_ssm_state back to input dtype.
- Mamba2Scan: same fp32 upcast for the multi-head SSM recurrence.
- GatedRMSNorm: upcast both hidden_states and gate to float32 for
  SiLU gating and variance computation, matching Zamba2RMSNormGated.
- PostGatedRMSNorm: use stash_type=1 for RMSNorm, fp32 for SiLU gate.
- GatedDeltaNet: upcast softplus/exp decay computation to float32.
- _RMSNorm (SSM internal): upcast variance computation to float32.

Update skills documentation:
- reusable-components: replace CastLike-only guidance with complete
  fp32 upcast pattern (upcast → compute → cast back), table of which
  ops need fp32, and when upcast is NOT needed.
- adding-a-new-model: rewrite troubleshooting #16 to cover both type
  mismatch and numerical drift, with the correct two-part pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 24, 2026
@gramalingam
Document dtype-agnostic CastLike pattern in skills
063db27 
Add troubleshooting entry #16 to adding-a-new-model skill covering the
hardcoded Cast(to=float32) pitfall that breaks bfloat16/float16 models.

Add design principle #7 (dtype-agnostic components) and a CastLike op
pattern section to the reusable-components skill, explaining when to use
CastLike vs Cast.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 24, 2026
@gramalingam
Upcast SSM recurrence and gated norms to fp32 for precision parity
d28a230 
Match HuggingFace's precision behaviour for numerically sensitive ops:
- SelectiveScan: upcast dt (softplus), A_log (exp), B, C, x, and state
  update to float32; cast output and new_ssm_state back to input dtype.
- Mamba2Scan: same fp32 upcast for the multi-head SSM recurrence.
- GatedRMSNorm: upcast both hidden_states and gate to float32 for
  SiLU gating and variance computation, matching Zamba2RMSNormGated.
- PostGatedRMSNorm: use stash_type=1 for RMSNorm, fp32 for SiLU gate.
- GatedDeltaNet: upcast softplus/exp decay computation to float32.
- _RMSNorm (SSM internal): upcast variance computation to float32.

Update skills documentation:
- reusable-components: replace CastLike-only guidance with complete
  fp32 upcast pattern (upcast → compute → cast back), table of which
  ops need fp32, and when upcast is NOT needed.
- adding-a-new-model: rewrite troubleshooting #16 to cover both type
  mismatch and numerical drift, with the correct two-part pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 27, 2026
@gramalingam
Document dtype-agnostic CastLike pattern in skills
3ef362c 
Add troubleshooting entry #16 to adding-a-new-model skill covering the
hardcoded Cast(to=float32) pitfall that breaks bfloat16/float16 models.

Add design principle #7 (dtype-agnostic components) and a CastLike op
pattern section to the reusable-components skill, explaining when to use
CastLike vs Cast.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 27, 2026
@gramalingam
Upcast SSM recurrence and gated norms to fp32 for precision parity
13c7281 
Match HuggingFace's precision behaviour for numerically sensitive ops:
- SelectiveScan: upcast dt (softplus), A_log (exp), B, C, x, and state
  update to float32; cast output and new_ssm_state back to input dtype.
- Mamba2Scan: same fp32 upcast for the multi-head SSM recurrence.
- GatedRMSNorm: upcast both hidden_states and gate to float32 for
  SiLU gating and variance computation, matching Zamba2RMSNormGated.
- PostGatedRMSNorm: use stash_type=1 for RMSNorm, fp32 for SiLU gate.
- GatedDeltaNet: upcast softplus/exp decay computation to float32.
- _RMSNorm (SSM internal): upcast variance computation to float32.

Update skills documentation:
- reusable-components: replace CastLike-only guidance with complete
  fp32 upcast pattern (upcast → compute → cast back), table of which
  ops need fp32, and when upcast is NOT needed.
- adding-a-new-model: rewrite troubleshooting #16 to cover both type
  mismatch and numerical drift, with the correct two-part pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 27, 2026
@gramalingam
Document dtype-agnostic CastLike pattern in skills
4ff811c 
Add troubleshooting entry #16 to adding-a-new-model skill covering the
hardcoded Cast(to=float32) pitfall that breaks bfloat16/float16 models.

Add design principle #7 (dtype-agnostic components) and a CastLike op
pattern section to the reusable-components skill, explaining when to use
CastLike vs Cast.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 27, 2026
@gramalingam
Upcast SSM recurrence and gated norms to fp32 for precision parity
c30ecca 
Match HuggingFace's precision behaviour for numerically sensitive ops:
- SelectiveScan: upcast dt (softplus), A_log (exp), B, C, x, and state
  update to float32; cast output and new_ssm_state back to input dtype.
- Mamba2Scan: same fp32 upcast for the multi-head SSM recurrence.
- GatedRMSNorm: upcast both hidden_states and gate to float32 for
  SiLU gating and variance computation, matching Zamba2RMSNormGated.
- PostGatedRMSNorm: use stash_type=1 for RMSNorm, fp32 for SiLU gate.
- GatedDeltaNet: upcast softplus/exp decay computation to float32.
- _RMSNorm (SSM internal): upcast variance computation to float32.

Update skills documentation:
- reusable-components: replace CastLike-only guidance with complete
  fp32 upcast pattern (upcast → compute → cast back), table of which
  ops need fp32, and when upcast is NOT needed.
- adding-a-new-model: rewrite troubleshooting #16 to cover both type
  mismatch and numerical drift, with the correct two-part pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 30, 2026
@gramalingam
Document dtype-agnostic CastLike pattern in skills
3410e58 
Add troubleshooting entry #16 to adding-a-new-model skill covering the
hardcoded Cast(to=float32) pitfall that breaks bfloat16/float16 models.

Add design principle #7 (dtype-agnostic components) and a CastLike op
pattern section to the reusable-components skill, explaining when to use
CastLike vs Cast.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
gramalingam added a commit that referenced this pull request Mar 30, 2026
@gramalingam
Upcast SSM recurrence and gated norms to fp32 for precision parity
2b9305d 
Match HuggingFace's precision behaviour for numerically sensitive ops:
- SelectiveScan: upcast dt (softplus), A_log (exp), B, C, x, and state
  update to float32; cast output and new_ssm_state back to input dtype.
- Mamba2Scan: same fp32 upcast for the multi-head SSM recurrence.
- GatedRMSNorm: upcast both hidden_states and gate to float32 for
  SiLU gating and variance computation, matching Zamba2RMSNormGated.
- PostGatedRMSNorm: use stash_type=1 for RMSNorm, fp32 for SiLU gate.
- GatedDeltaNet: upcast softplus/exp decay computation to float32.
- _RMSNorm (SSM internal): upcast variance computation to float32.

Update skills documentation:
- reusable-components: replace CastLike-only guidance with complete
  fp32 upcast pattern (upcast → compute → cast back), table of which
  ops need fp32, and when upcast is NOT needed.
- adding-a-new-model: rewrite troubleshooting #16 to cover both type
  mismatch and numerical drift, with the correct two-part pattern.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: G Ramalingam <grama@microsoft.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@justinchuby