feat: refactor GpuType enum names, add B200 and RTX PRO 6000s

[KAJdev]

Adds GpuType.RTX_PRO_6000_BLACKWELL_SERVER_EDITION, GpuType.RTX_PRO_6000_BLACKWELL_WORKSTATION_EDITION, GpuType.RTX_PRO_6000_BLACKWELL_MAX_Q_WORKSTATION_EDITION, and GpuType.B200.

Adds GpuGroup.BLACKWELL_180 and GpuGroup.BLACKWELL_96 enum members

Also refactors the GpuType enum to remove the NVIDIA_ prefix from all enum identifiers.

[runpod-Henrik]

QA Report

Status: MERGE WITH NOTES
PR: #233 — feat!: refactor GpuType enum names, add B200 and RTX PRO 6000s
Agent: flash-qa (PR mode)

Breaking Change Analysis

Renamed enums (NVIDIA_ prefix removed from all 16 members):

Old Name	New Name
NVIDIA_GEFORCE_RTX_4090	GEFORCE_RTX_4090
NVIDIA_GEFORCE_RTX_5090	GEFORCE_RTX_5090
NVIDIA_RTX_6000_ADA_GENERATION	RTX_6000_ADA_GENERATION
NVIDIA_H100_80GB_HBM3	H100_80GB_HBM3
NVIDIA_RTX_A4000	RTX_A4000
NVIDIA_RTX_A4500	RTX_A4500
NVIDIA_RTX_4000_ADA_GENERATION	RTX_4000_ADA_GENERATION
NVIDIA_RTX_2000_ADA_GENERATION	RTX_2000_ADA_GENERATION
NVIDIA_RTX_A5000	RTX_A5000
NVIDIA_L4	L4
NVIDIA_GEFORCE_RTX_3090	GEFORCE_RTX_3090
NVIDIA_A40	A40
NVIDIA_RTX_A6000	RTX_A6000
NVIDIA_A100_80GB_PCIe	A100_80GB_PCIe
NVIDIA_A100_SXM4_80GB	A100_SXM4_80GB
NVIDIA_H200	H200

Backward compatibility: No aliases or deprecation warnings provided. Old enum names will raise AttributeError immediately.

Downstream impact:

• flash-examples: 1 file uses old names (01_getting_started/05_gpu_selection/gpu_worker.py uses GpuType.NVIDIA_GEFORCE_RTX_4090) -- will break on upgrade
• flash-worker: Does not reference GpuType enum member names directly -- no impact
• Any user code referencing GpuType.NVIDIA_* will break

CI Results

All checks passing:

• Quality Gates: 3.10, 3.11, 3.12, 3.13, 3.14 -- all pass
• Build Package -- pass

New GPU Validation

GPU	Enum Name	API ID (value)	In POOLS_TO_TYPES?
B200	B200	NVIDIA B200	Yes (BLACKWELL_180)
RTX PRO 6000 Server	RTX_PRO_6000_BLACKWELL_SERVER_EDITION	NVIDIA RTX PRO 6000 Blackwell Server Edition	No
RTX PRO 6000 Workstation	RTX_PRO_6000_BLACKWELL_WORKSTATION_EDITION	NVIDIA RTX PRO 6000 Blackwell Workstation Edition	No
BLACKWELL_180 (group)	N/A	BLACKWELL_180	N/A (is the pool)

Issues Found

[P1] RTX PRO 6000 variants missing from POOLS_TO_TYPES
Both RTX_PRO_6000_BLACKWELL_SERVER_EDITION and RTX_PRO_6000_BLACKWELL_WORKSTATION_EDITION are not mapped to any GpuGroup in POOLS_TO_TYPES. This means _pool_from_gpu_type() returns None for them, and GpuGroup.to_gpu_ids_str() will add None to the pool_ids set, producing malformed API strings. Users who specify these GPUs directly will hit a bug at deploy time.

Suggested fix: either add them to an existing pool (e.g., ADA_48_PRO or a new Blackwell pool) or add a guard in _pool_from_gpu_type() / to_gpu_ids_str().

[P2] No backward-compatible aliases for renamed enums
Since this is published on PyPI as a public SDK, users on runpod-flash<1.4.2 who upgrade will get immediate AttributeError with no deprecation path. Consider adding aliases with DeprecationWarning in GpuType.__init_subclass__ or as class attributes:

# Backward-compatible aliases (deprecated)
NVIDIA_GEFORCE_RTX_4090 = GEFORCE_RTX_4090
NVIDIA_L4 = L4
# etc.

[P3] No test coverage for new GPU types
The test file (test_gpu_ids.py) does not test B200, RTX_PRO_6000_*, or BLACKWELL_180. No round-trip test for to_gpu_ids_str / from_gpu_ids_str with new types.

PR Diff Analysis

• No bare exceptions
• Existing enum values (API strings) unchanged -- only member names renamed
• Breaking changes documented in commit message (feat!:)
• Skeleton template README updated with new names
• RTX PRO 6000 variants not in any pool mapping -- potential runtime bug
• No deprecation aliases for old names
• flash-examples repo needs coordinated update

Recommendation

MERGE WITH NOTES -- The core rename is clean and CI passes. However, before or shortly after merge:

1. Must fix the RTX PRO 6000 pool mapping gap (P1) -- these GPUs will produce broken API calls if used
2. Should add deprecation aliases for the old NVIDIA_* names to avoid a hard break for existing users (P2)
3. Should add tests for new GPU types and round-trip pool serialization (P3)
4. Must coordinate a flash-examples update for gpu_worker.py

---

Generated by flash-qa agent